Datasets:

yxzwayne

/

abs

like 0

2402.11960v1.pdf

DB-LLM: Accurate Dual-Binarization for Efficient LLMs Hong Chen1*, Chengtao Lv1*, Liang Ding2, Haotong Qin1, Xiabin Zhou4, Yifu Ding1, Xuebo Liu3, Min Zhang3, Jinyang Guo1, Xianglong Liu1†, Dacheng Tao2 1Beihang University2The University of Sydney 3Harbin Institute of Technology, Shenzhen4Jiangsu University {18373205, lvchengtao, qinhaotong, xlliu}@buaa.edu.cn ,liangding.liam@gmail.com Abstract Large language models (LLMs) have signifi- cantly advanced the field of natural language processing, while the expensive memory and computation consumption impede their practi- cal deployment. Quantization emerges as one of the most effective methods for improving the computational efficiency of LLMs. How- ever, existing ultra-low-bit quantization always causes severe accuracy drops. In this paper, we empirically relieve the micro and macro characteristics of ultra-low bit quantization and present a novel Dual-Binarization method for LLM s, namely DB-LLM . For the micro-level, we take both the accuracy advantage of 2-bit- width and the efficiency advantage of binariza- tion into account, introducing Flexible Dual Bi- narization (FDB ). By splitting 2-bit quantized weights into two independent sets of binaries, FDB ensures the accuracy of representations and introduces flexibility, utilizing the efficient bitwise operations of binarization while retain- ing the inherent high sparsity of ultra-low bit quantization. For the macro-level, we find the distortion that exists in the prediction of LLM after quantization, which is specified as the de- viations related to the ambiguity of samples. We propose the Deviation-Aware Distillation (DAD ) method, enabling the model to focus differently on various samples. Comprehensive experiments show that our DB-LLM not only significantly surpasses the current State-of-The- Art (SoTA) in ultra-low bit quantization ( e.g., perplexity decreased from 9.64 to 7.23), but also achieves an additional 20% reduction in computational consumption compared to the SOTA method under the same bit-width. Our code will be released soon. 1 Introduction Recently, Large Language Models (LLMs), such as ChatGPT (Brown et al., 2020) and LLaMA (Tou- vron et al., 2023a) have catalyzed a paradigm shift *Equal contribution. †Corresponding author. 2 4 8 16 32 64 128 Model Size (GB, log scale)5101520Perplexity 7.596.535.524.84FP16 AWQ 3bit GPTQ 2bit DB-LLM (Ours)Figure 1: The perplexity on WikiText2 for LLaMA family models. 2-bit DB-LLM is close to FP results and surpasses 3-bit AWQ by a large margin. in Natural Language Processing (NLP), marking a significant milestone in the AI revolution. Their unprecedented capabilities evolved from a mas- sive memory footprint ( e.g., billion-scale parame- ters), which constrains the widespread application of LLMs on resource-limited devices. Several com- pression schemes are thus proposed to reduce the memory demands of LLMs, which can be roughly categorized into weight quantization (Frantar et al., 2022; Lin et al., 2023), network pruning (Sun et al., 2023; Ma et al., 2023; He et al., 2022), knowledge distillation (Gu et al., 2023; Zhong et al., 2024) and low-rank factorization (Xu et al., 2023; Yuan et al., 2023). Among these methods, weight quantization is highly effective and practical since it achieves the best trade-off between the performance and the cost of the compression process. Nevertheless, although many works (Shao et al., 2023; Shang et al., 2023) attempt to quantize LLMs to ultra-low-bit ( e.g., 2- bit), their performance is unsatisfactory and falls far short of industrial application requirements. Ultra-low-bit quantization ( ≤4 bits), as an ex- tremely efficient form of quantization, enjoys over 8×memory compression ratio. Despite these spe- cialized weight-only quantization schemes achiev- ing savings in storage consumption, they still can-arXiv:2402.11960v1 [cs.LG] 19 Feb 2024

2210.13382.pdf

Published as a conference paper at ICLR 2023 EMERGENT WORLD REPRESENTATIONS : EXPLORING A SEQUENCE MODEL TRAINED ON A SYNTHETIC TASK Kenneth Li∗ Harvard UniversityAspen K. Hopkins Massachusetts Institute of TechnologyDavid Bau Northeastern University Fernanda Vi ´egas Harvard UniversityHanspeter Pfister Harvard UniversityMartin Wattenberg Harvard University ABSTRACT Language models show a surprising range of capabilities, but the source of their apparent competence is unclear. Do these networks just memorize a collection of surface statistics, or do they rely on internal representations of the process that generates the sequences they see? We investigate this question in a synthetic setting by applying a variant of the GPT model to the task of predicting legal moves in a simple board game, Othello. Although the network has no a priori knowledge of the game or its rules, we uncover evidence of an emergent nonlinear internal representation of the board state. Interventional experiments indicate this representation can be used to control the output of the network. By leveraging these intervention techniques, we produce “latent saliency maps” that help explain predictions.1 1 I NTRODUCTION Recent language models have shown an intriguing range of capabilities. Networks trained on a simple “next-word” prediction task are apparently capable of many other things, such as solving logic puzzles or writing basic code.2Yet how this type of performance emerges from sequence predictions remains a subject of current debate. Some have suggested that training on a sequence modeling task is inherently limiting. The arguments range from philosophical (Bender & Koller, 2020) to mathematical (Merrill et al., 2021). A common theme is that seemingly good performance might result from memorizing “surface statistics,” i.e., a long list of correlations that do not reflect a causal model of the process generating the sequence. This issue is of practical concern, since relying on spurious correlations may lead to problems on out-of-distribution data (Bender et al., 2021; Floridi & Chiriatti, 2020). On the other hand, some tantalizing clues suggest language models may do more than collect spurious correlations, instead building interpretable world models —that is, understandable models of the process producing the sequences they are trained on. Recent evidence suggests language models can develop internal representations for very simple concepts, such as color, direction Abdou et al. (2021); Patel & Pavlick (2022), or tracking boolean states during synthetic tasks (Li et al., 2021) (see Related Work (section 6) for more detail). A promising approach to studying the emergence of world models is used by Toshniwal et al. (2021), which explores language models trained on chess move sequences. The idea is to analyze the behavior of a standard language modeling architecture in a well-understood, constrained setting. The paper finds that these models learn to predict legal chess moves with high accuracy. Furthermore, by analyzing predicted moves, the paper shows that the model appears to track the board state. The authors stop short, however, of exploring the form of any internal representations. Such an ∗Correspondence to keli@g.harvard.edu 1Codes at https://github.com/likenneth/othello_world 2See Srivastava et al. (2022) for an encyclopedic list of examples. 1arXiv:2210.13382v4 [cs.LG] 27 Feb 2023

1809.04281.pdf

MUSIC TRANSFORMER : GENERATING MUSIC WITH LONG -TERM STRUCTURE Cheng-Zhi Anna Huang∗Ashish Vaswani Jakob Uszkoreit Noam Shazeer Ian Simon Curtis Hawthorne Andrew M. Dai Matthew D. Hoffman Monica Dinculescu Douglas Eck Google Brain ABSTRACT Music relies heavily on repetition to build structure and meaning. Self-reference occurs on multiple timescales, from motifs to phrases to reusing of entire sections of music, such as in pieces with ABA structure. The Transformer (Vaswani et al., 2017), a sequence model based on self-attention, has achieved compelling results in many generation tasks that require maintaining long-range coherence. This suggests that self-attention might also be well-suited to modeling music. In musical composition and performance, however, relative timing is critically important. Existing approaches for representing relative positional information in the Transformer modulate attention based on pairwise distance (Shaw et al., 2018). This is impractical for long sequences such as musical compositions since their memory complexity for intermediate relative information is quadratic in the sequence length. We propose an algorithm that reduces their intermediate memory requirement to linear in the sequence length. This enables us to demonstrate that a Transformer with our modified relative attention mechanism can generate minute- long compositions (thousands of steps, four times the length modeled in Oore et al. (2018)) with compelling structure, generate continuations that coherently elaborate on a given motif, and in a seq2seq setup generate accompaniments conditioned on melodies1. We evaluate the Transformer with our relative attention mechanism on two datasets, JSB Chorales and Piano-e-Competition, and obtain state-of-the-art results on the latter. 1 I NTRODUCTION A musical piece often consists of recurring elements at various levels, from motifs to phrases to sections such as verse-chorus. To generate a coherent piece, a model needs to reference elements that came before, sometimes in the distant past, repeating, varying, and further developing them to create contrast and surprise. Intuitively, self-attention (Parikh et al., 2016) appears to be a good match for this task. Self-attention over its own previous outputs allows an autoregressive model to access any part of the previously generated output at every step of generation. By contrast, recurrent neural networks have to learn to proactively store elements to be referenced in a fixed size state or memory, potentially making training much more difficult. We believe that repeating self-attention in multiple, successive layers of a Transformer decoder (Vaswani et al., 2017) helps capture the multiple levels at which self-referential phenomena exist in music. In its original formulation, the Transformer relies on absolute position representations, using either positional sinusoids or learned position embeddings that are added to the per-position input repre- sentations. Recurrent and convolutional neural networks instead model position in relative terms: RNNs through their recurrence over the positions in their input, and CNNs by applying kernels that effectively choose which parameters to apply based on the relative position of the covered input representations. ∗Google AI Resident. Correspondence to: Cheng-Zhi Anna Huang <annahuang@google.com> 1Samples are available for listening at https://storage.googleapis.com/music-transformer/index.html 1arXiv:1809.04281v3 [cs.LG] 12 Dec 2018

NeurIPS-2022-training-language-models-to-follow-instructions-with-human-feedback-Paper-Conference.pdf

Training language models to follow instructions with human feedback Long Ouyang∗Jeff Wu∗Xu Jiang∗Diogo Almeida∗Carroll L. Wainwright∗ Pamela Mishkin∗Chong Zhang Sandhini Agarwal Katarina Slama Alex Ray John Schulman Jacob Hilton Fraser Kelton Luke Miller Maddie Simens Amanda Askell†Peter Welinder Paul Christiano∗† Jan Leike∗Ryan Lowe∗ OpenAI Abstract Making language models bigger does not inherently make them better at following a user’s intent. For example, large language models can generate outputs that are untruthful, toxic, or simply not helpful to the user. In other words, these models are notaligned with their users. In this paper, we show an avenue for aligning language models with user intent on a wide range of tasks by fine-tuning with human feedback. Starting with a set of labeler-written prompts and prompts submitted through a language model API, we collect a dataset of labeler demonstrations of the desired model behavior, which we use to fine-tune GPT-3 using supervised learning. We then collect a dataset of rankings of model outputs, which we use to further fine-tune this supervised model using reinforcement learning from human feedback. We call the resulting models InstructGPT . In human evaluations on our prompt distribution, outputs from the 1.3B parameter InstructGPT model are preferred to outputs from the 175B GPT-3, despite having 100x fewer parameters. Moreover, InstructGPT models show improvements in truthfulness and reductions in toxic output generation while having minimal performance regressions on public NLP datasets. Even though InstructGPT still makes simple mistakes, our results show that fine-tuning with human feedback is a promising direction for aligning language models with human intent. 1 Introduction Large language models (LMs) can be prompted to perform a range of natural language process- ing (NLP) tasks, given some examples of the task as input. However, these models often express unintended behaviors such as making up facts, generating biased or toxic text, or simply not following user instructions (Bender et al., 2021; Bommasani et al., 2021; Kenton et al., 2021; Weidinger et al., 2021; Tamkin et al., 2021; Gehman et al., 2020). This is because the language modeling objective ∗Primary authors. This was a joint project of the OpenAI Alignment team. RL and JL are the team leads. Corresponding author: lowe@openai.com . †Work done while at OpenAI. Current affiliations: AA: Anthropic; PC: Alignment Research Center. 36th Conference on Neural Information Processing Systems (NeurIPS 2022).

2305.12132.pdf

Can Public Large Language Models Help Private Cross-device Federated Learning? Boxin Wang3∗, Yibo Jacky Zhang4, Yuan Cao2, Bo Li3, H. Brendan McMahan1, Sewoong Oh1, Zheng Xu1, Manzil Zaheer2 1Google Research,2Google Deepmind,3UIUC,4Stanford Abstract We study (differentially) private federated learning (FL) of language models. The lan- guage models in cross-device FL are relatively small, which can be trained with meaning- ful formal user-level differential privacy (DP) guarantees when massive parallelism in train- ing is enabled by the participation of a mod- erate size of users. Recently, public data has been used to improve privacy-utility trade-offs for both large and small language models. In this work, we provide a systematic study of us- ing large-scale public data and LLMs to help differentially private training of on-device FL models, and further improve the privacy-utility tradeoff by techniques of distillation. More- over, we propose a novel distribution match- ing algorithm with theoretical grounding to sample public data close to private data distri- bution, which significantly improves the sam- ple efficiency of (pre-)training on public data. The proposed method is efficient and effective for training private model by taking advantage of public data, especially for customized on- device architectures that do not have ready-to- use pre-trained models. 1 Introduction Federated Learning (FL) (McMahan et al., 2017, 2018; Kairouz et al., 2019) is designed to collabo- ratively train a global model on decentralized data across user clients while protecting data privacy. FL emerged as an effective privacy-preserving so- lution of training (language) models, as rich text data are generated by users, which may contain sen- sitive and personal information. After McMahan et al. (2017) proposed to train on-device recurrent neural network models, FL has been widely used in various natural language processing applications and products, including next-word prediction (Hard ∗Part of the work was done while Boxin Wang was an intern at Google. Correspondence to: Boxin Wang boxinw2@illinois.edu and Zheng Xu xuzheng@google.com .et al., 2018), keyword spotting (Hard et al., 2020), and out-of-vocabulary word discovery (Chen et al., 2019). To further protect user privacy, Differential Pri- vacy (DP) (Dwork et al., 2006; Dwork, 2011; Dwork and Roth, 2014; McMahan et al., 2018) is introduced to provide formal privacy guarantees of models trained by federated learning. DP for deep learning explicitly adds random noise with bounded sensitivity to a training process ( e.g., DP- SGD (Abadi et al., 2016)), ensuring a quantifiable similarity in output model distributions when the training dataset changes. When combining DP with FL, a variant of DP-SGD called DP-FedAvg (McMahan et al., 2018)) is applied to guarantee user-level DP (Dwork, 2010). Current research pri- marily focuses on applying user-level DP to small on-device models with fewer than 10 million pa- rameters (McMahan et al., 2018; Kairouz et al., 2021; Ramaswamy et al., 2020). The model size is limited due to challenges such as significant DP noise required to preserve privacy (Li et al., 2021) and the communication costs in cross-device FL. Recent advances in large language models (LLMs) (Thoppilan et al., 2022; Radford et al., 2019; Brown et al., 2020; Devlin et al., 2019; Raffel et al., 2020) have revolutionized natural language processing (NLP) and achieved unprecedented per- formance on various tasks such as text generation, machine translation, and sentiment analysis. How- ever, their success comes at a cost of requiring mas- sive amounts of computational resources, making them difficult to deploy on resource-constrained devices such as smartphones, tablets, or other edge devices. Additionally, there are concerns regarding the user privacy in various aspects such as memoriz- ing personal information in training, and exposing private query in inference. Recent work explore incorporating public infor- mation to improve privacy-utility trade-off in ap- plying DP for (large) LMs (Yu et al., 2022; Li et al.,arXiv:2305.12132v1 [cs.LG] 20 May 2023

2201.02867v3.pdf

Deep Generative Modeling for Volume Reconstruction in Cryo-Electron Microscopy Claire Donnat1+, Axel Levy2,3, Fr´ed´eric Poitevin3, Ellen Zhong4, and Nina Miolane5*+ 1University of Chicago, Department of Statistics, Chicago, Illinois, USA 2Stanford University, Department of Electrical Engineering, Stanford, CA, USA 3LCLS, SLAC National Accelerator Laboratory, Menlo Park, CA, USA 4Massachusetts Institute of Technology, Computer Science and Artificial Intelligence Lab, Boston, MA, USA 5University of California Santa Barbara, Department of Electrical & Computer Engineering, Santa Barbara, CA, USA *ninamiolane@ucsb.edu +these authors contributed equally to this work ABSTRACT Recent breakthroughs in high-resolution imaging of biomolecules in solution with cryo-electron microscopy (cryo-EM) have unlocked new doors for the reconstruction of molecular volumes, thereby promising further advances in biology, chemistry, and pharmacological research. Recent next-generation volume reconstruction algorithms that combine generative modeling with end-to-end unsupervised deep learning techniques have shown promising preliminary results, but still face considerable technical and theoretical hurdles when applied to experimental cryo-EM images. In light of the proliferation of such methods, we propose here a critical review of recent advances in the field of deep generative modeling for cryo-EM volume reconstruction . The present review aims to (i) unify and compare these new methods using a consistent statistical framework, (ii) present them using a terminology familiar to machine learning researchers and computational biologists with no specific background in cryo-EM, and (iii) provide the necessary perspective on current advances to highlight their relative strengths and weaknesses, along with outstanding bottlenecks and avenues for improvements in the field. This review might also raise the interest of computer vision practitioners, as it highlights significant limits of deep generative models in low signal-to-noise regimes — therefore emphasizing a need for new theoretical and methodological developments. Introduction Electron beam Particles: biomolecules “flash frozen” in solution 2D projections Figure 1. Acquisition of 2D cryo-EM images (2D projections) from 3D biomolecular volumes.High-resolution reconstruction of molecular volumes from single par- ticle images has the potential to facilitate new breakthroughs in our ability to understand fundamental biological mechanisms and engineer macromolecular function7, 52. In this context, cryo-electron microscopy (cryo-EM) has fostered a revolution in structural biology by allowing the imaging of biomolecules in solution at atomic resolution8, 9. However, the estimation of these molecules’ 3-dimensional (3D) volume from cryo-EM data continues to pose a formidable challenge. In this setting, observations are limited to the raw 2D projections of molecules (also called particles) relative to an incoming electron beam, while their 3D ori- entation and position (jointly called poses) are unknown — see Figure 1. Reconstructing molecular volumes therefore also requires recovering a number of hidden variables such as each particle’s 3D orientation. The difficulty of this task is further compounded by a combination of factors, including the variability in the shape of any given molecule (also referred to as structural “heterogeneity”), the non-linear physics of the data acquisition process, as well as extremely low signal-to-noise ratios — concepts formalized in the image formation model below. Image Formation Model. The process of image formation in cryo-EM involves several physical phenomena, including pairwise interactions between atoms, interactions between the electron beam and the molecule’s electrostatic potential, or microscope effects. We refer the reader to Dill et al.10, Kohl and Reimer11, and Vulovic et al.12for in-depth descriptions of these phenomena. Nonetheless, in most cases12, 13, each image Xiin a dataset of nimages of single particles can be modeled as a random sample from the following generative model: Xi=PSF i∗(ti◦Π2D◦Ri)(V(i))+εi, with i=1···n. (1)arXiv:2201.02867v3 [eess.IV] 26 May 2022

2304.02034.pdf

Effective Theory of Transformers at Initialization Emily Dinan,∗Sho Yaida,†and Susan Zhang‡ Meta AI Meta Platforms, Inc.§ We perform an effective-theory analysis of forward–backward signal propagation in wide and deep Transformers, i.e., residual neural networks with multi-head self-attention blocks and multilayer perceptron blocks. This analysis suggests particular width scalings of initial- ization and training hyperparameters for these models. We then take up such suggestions, training Vision and Language Transformers in practical setups. ∗Electronic address: edinan@meta.com †Electronic address: shoyaida@meta.com ‡Electronic address: susanz@meta.com §The author ordering was determined by the hypothetical coin toss that 100%-respects the alphabetical ordering.arXiv:2304.02034v1 [cs.LG] 4 Apr 2023

2307.12950.pdf

RLCD: R EINFORCEMENT LEARNING FROM CONTRAST DISTILLATION FOR LANGUAGE MODEL ALIGNMENT Kevin Yang1,2Dan Klein1Asli Celikyilmaz2Nanyun Peng3Yuandong Tian2 1UC Berkeley,2Meta AI,3UCLA {yangk,klein}@berkeley.edu,{aslic,yuandong}@meta.com,violetpeng@cs.ucla.edu ABSTRACT We propose Reinforcement Learning from Contrast Distillation (RLCD), a method for aligning language models to follow natural language principles without using human feedback. RLCD trains a preference model using simulated preference pairs that contain both a high-quality and low-quality example, generated using contrasting positive and negative prompts. The preference model is then used to improve a base unaligned language model via reinforcement learning. Empirically, RLCD outperforms RLAIF (Bai et al., 2022b) and context distillation (Huang et al., 2022) baselines across three diverse alignment tasks—harmlessness, helpfulness, and story outline generation—and on both 7B and 30B model scales for preference data simulation. 1 I NTRODUCTION Reinforcement Learning from Human Feedback (RLHF) has recently been used to great effect to align pretrained large language models (LLMs) to human preferences, optimizing for desirable qualities like harmlessness and helpfulness (Bai et al., 2022a) and achieving state-of-the-art results across a variety of natural language tasks (OpenAI, 2023). A standard RLHF procedure fine-tunes an initial unaligned LLM using an RL algorithm such as PPO (Schulman et al., 2017), optimizing the LLM to align with human preferences. RLHF is thus critically dependent on a reward model derived from human-labeled preferences, typically pairwise preferences on LLM outputs (o1, o2)generated from a shared prompt p. However, collecting human pairwise preference data, especially high-quality data, may be expensive and time consuming at scale. To address this problem, approaches have been proposed to obtain labels without human annotation, such as Reinforcement Learning from AI Feedback (RLAIF) and context distillation. RLAIF approaches (e.g., Bai et al. (2022b)) simulate human pairwise preferences by scoring o1and o2with an LLM (Figure 1 center); the scoring LLM is often the same as the one used to generate the original pairs (o1, o2). Of course, the resulting LLM pairwise preferences will be somewhat noisier compared to human labels. However, this problem is exacerbated by using the same prompt pto generate both o1ando2, causing o1ando2to often be of very similar quality and thus hard to differentiate (e.g., Table 1). Consequently, training signal can be overwhelmed by label noise, yielding lower-quality preference data. Meanwhile, context distillation methods (e.g., Sun et al. (2023)) create more training signal by modifying the initial prompt p. The modified prompt p+typically contains additional context encouraging a directional attribute change in the output o+(Figure 1 right). However, context distillation methods only generate a single output o+per prompt p+, which is then used for supervised fine-tuning, losing the pairwise preferences which help RLHF-style approaches to derive signal from the contrast between outputs. Multiple works have observed that RL approaches using preference models for pairwise preferences can substantially improve over supervised fine-tuning by itself when aligning LLMs (Ouyang et al., 2022; Dubois et al., 2023). Therefore, while both RLAIF and context distillation approaches have already been successfully applied in practice to align language models, we posit that it may be even more effective to combine 1arXiv:2307.12950v1 [cs.CL] 24 Jul 2023

2206.14486.pdf

Beyond neural scaling laws: beating power law scaling via data pruning Ben Sorscher∗ ∗1Robert Geirhos∗2Shashank Shekhar3 Surya Ganguli1,3§Ari S. Morcos3§ ∗equal contribution 1Department of Applied Physics, Stanford University 2University of Tübingen 3Meta AI (FAIR) §Joint senior authors Abstract Widely observed neural scaling laws, in which error falls off as a power of the training set size, model size, or both, have driven substantial performance im- provements in deep learning. However, these improvements through scaling alone require considerable costs in compute and energy. Here we focus on the scaling of error with dataset size and show how in theory we can break beyond power law scaling and potentially even reduce it to exponential scaling instead if we have access to a high-quality data pruning metric that ranks the order in which training examples should be discarded to achieve any pruned dataset size. We then test this improved scaling prediction with pruned dataset size empirically, and indeed observe better than power law scaling in practice on ResNets trained on CIFAR-10, SVHN, and ImageNet. Next, given the importance of finding high-quality pruning metrics, we perform the first large-scale benchmarking study of ten different data pruning metrics on ImageNet. We find most existing high performing metrics scale poorly to ImageNet, while the best are computationally intensive and require labels for every image. We therefore developed a new simple, cheap and scalable self-supervised pruning metric that demonstrates comparable performance to the best supervised metrics. Overall, our work suggests that the discovery of good data-pruning metrics may provide a viable path forward to substantially improved neural scaling laws, thereby reducing the resource costs of modern deep learning. 1 Introduction Empirically observed neural scaling laws [ 1,2,3,4,5,6,7,8] in many domains of machine learning, including vision, language, and speech, demonstrate that test error often falls off as a power law with either the amount of training data, model size, or compute. Such power law scaling has motivated significant societal investments in data collection, compute, and associated energy consumption. However, power law scaling is extremely weak and unsustainable. For example, a drop in error ∗work done during an internship at Meta AI (FAIR) 36th Conference on Neural Information Processing Systems (NeurIPS 2022).arXiv:2206.14486v6 [cs.LG] 21 Apr 2023

2305.16381.pdf

DPOK: Reinforcement Learning for Fine-tuning Text-to-Image Diffusion Models Ying Fan˚,1,2, Olivia Watkins3, Yuqing Du3, Hao Liu3, Moonkyung Ryu1, Craig Boutilier1, Pieter Abbeel3,Mohammad Ghavamzadeh1,Kangwook Lee2,Kimin Lee˚,1 ˚Equal technical contribution 1Google Research2University of Wisconsin-Madison3UC Berkeley Abstract Learning from human feedback has been shown to improve text-to-image models. These techniques first learn a reward function that captures what humans care about in the task and then improve the models based on the learned reward function. Even though relatively simple approaches (e.g., rejection sampling based on reward scores) have been investigated, fine-tuning text-to-image models with the reward function remains challenging. In this work, we propose using online reinforcement learning (RL) to fine-tune text-to-image models. We focus on diffusion models , defining the fine-tuning task as an RL problem, and updating the pre-trained text-to-image diffusion models using policy gradient to maximize the feedback- trained reward. Our approach, coined DPOK, integrates policy optimization with KL regularization. We conduct an analysis of KL regularization for both RL fine-tuning and supervised fine-tuning. In our experiments, we show that DPOK is generally superior to supervised fine-tuning with respect to both image-text alignment and image quality. 1 Introduction Recent advances in diffusion models [10,36,37], together with pre-trained text encoders (e.g., CLIP [ 28], T5 [ 29]) have led to impressive results in text-to-image generation. Large-scale text-to- image models, such as Imagen [ 33], Dalle-2 [ 30] and Stable Diffusion [ 31], generate high-quality, creative images given novel text prompts. However, despite these advances, current models have systematic weaknesses. For example, current models have a limited ability to compose multiple objects [ 6,7,26]. They also frequently encounter difficulties when generating objects with specified colors and counts [12, 18]. Learning from human feedback (LHF) has proven to be an effective means to overcome these limitations [ 14,18,41,43]. Lee et al. [18] demonstrate that certain properties, such as generating objects with specific colors, counts, and backgrounds, can be improved by learning a reward function from human feedback, followed by fine-tuning the text-to-image model using supervised learning. They show that simple supervised fine-tuning based on reward-weighted loss can improve the reward scores, leading to better image-text alignment. However, supervised fine-tuning often induces a deterioration in image quality (e.g., over-saturated or non-photorealistic images). This is likely due to the model being fine-tuned on a fixed dataset that is generated by a pre-trained model (Figure 1(a)). In this work, we explore using online reinforcement learning (RL) for fine-tuning text-to-image diffusion models (Figure 1(b)). We show that optimizing the expected reward of a diffusion model’s image output is equivalent to performing policy gradient on a multi-step diffusion model under certain regularity assumptions. We also incorporate Kullback–Leibler (KL) divergence with respect to the pre-trained model as regularization in an online manner, treating this as an implicit reward. Preprint. Under review.arXiv:2305.16381v1 [cs.LG] 25 May 2023

10.1016.j.cell.2024.01.036.pdf

Article Structure of the plant plastid-encoded RNA polymerase Graphical abstract Highlights dStructure of the chloroplast transcription complex dFifteen nuclear-encoded subunits encase the plastid- encoded polymerase dSubunits PAP1 and PAP2 interact with the DNA and themRNA, respectively dStructure-guided insights into enzymatic activities ofsubunitsAuthors A´ngel Vergara-Cruces, Ishika Pramanick, David Pearce, Vinod K. Vogirala,Matthew J. Byrne, Jason K.K. Low,Michael W. Webster Correspondence michael.webster@jic.ac.uk In brief Structural characterization of thechloroplast RNA polymerase thattranscribes photosynthetic genesprovides insight into its composition,assembly, and evolution. Vergara-Cruces et al., 2024, Cell 187, 1145–1159 February 29, 2024 Crown Copyright ª2024 Published by Elsevier Inc. https://doi.org/10.1016/j.cell.2024.01.036 ll

99_on_recovering_higher_order_int.pdf

ONRECOVERING HIGHER -ORDER INTERACTIONS FROM PROTEIN LANGUAGE MODELS Darin Tsui & Amirali Aghazadeh School of Electrical and Computer Engineering Georgia Institute of Technology Atlanta, GA 30332, USA {darint,amiralia }@gatech.edu ABSTRACT Protein language models leverage evolutionary information to perform state-of- the-art 3D structure and zero-shot variant prediction. Yet, extracting and explain- ingallthe mutational interactions that govern model predictions remains diffi- cult as it requires querying the entire amino acid space for nsites using 20nse- quences, which is computationally expensive even for moderate values of n(e.g., n∼10). Although approaches to lower the sample complexity exist, they of- ten limit the interpretability of the model to just single and pairwise interactions. Recently, computationally scalable algorithms relying on the assumption of spar- sity in the Fourier domain have emerged to learn interactions from experimental data. However, extracting interactions from language models poses unique chal- lenges: it’s unclear if sparsity is always present or if it is the only metric needed to assess the utility of Fourier algorithms. Herein, we develop a framework to do a systematic Fourier analysis of the protein language model ESM2 applied on three proteins—green fluorescent protein (GFP), tumor protein P53 (TP53), and G domain B1 (GB1)—across various sites for 228 experiments. We demonstrate that ESM2 is dominated by three regions in the sparsity-ruggedness plane, two of which are better suited for sparse Fourier transforms. Validations on two sam- ple proteins demonstrate recovery of all interactions with R2= 0.72in the more sparse region and R2= 0.66in the more dense region, using only 7 million out of2010∼1013ESM2 samples, reducing the computational time by a stagger- ing factor of 15,000. All codes and data are available on our GitHub repository https://github.com/amirgroup-codes/InteractionRecovery. 1 I NTRODUCTION Recent advances in transformer-based deep learning models have leveraged evolutionary informa- tion to learn biological patterns in protein sequences. These models, encompassing up to 15 billion learnable parameters, are trained on amino acid sequences stored in databases such as UniProt (Lin et al., 2023; Consortium, 2015). In particular, masked language models have been demonstrated to achieve state-of-the-art performance in zero-shot variant effect and protein structure prediction without the need for explicit training (Meier et al., 2021; Brandes et al., 2023). Hence, it’s widely believed that protein language models encapsulate representations that reflect the fundamental rules of biology and physics (Rives et al., 2021; Rao et al., 2020). However, further applications of protein language models, e.g., for knowledge discovery, are hindered due to the challenge of interpreting the biological interactions that underlie their predictions. In principle, if we wanted to learn the structural impact of variants underlying nmutational sites in a protein, referred to as the region’s landscape, we could query these language models on all possible 20nmutational combinations (for all 20 standard amino acids). However, computational challenges would make such an endeavor nearly unrunnable at a large scale. For instance, on four NVIDIA RTX A6000s, each sample takes about 0.01 seconds to compute. It would take 20n×0.01 = 32000 seconds, or around nine hours, to compute all possible combinations for n= 5. However, even just increasing the length to n= 8would make the entire space take 194 years to complete. 1

langegabelriedmiller2011chapter.pdf

Batch Reinforcement Learning Sascha Lange, Thomas Gabel, and Martin Riedmiller Abstract Batch reinforcement learning is a subfield of dynamic programming-based reinforcement learning. Originally defined as the task of learning the best possible policy from a fixed set of a priori-known transition samples, the (batch) algorithms developed in this field can be easily adapted to the classical online case, where the agent interacts with the environment while learning. Due to the efficient use of col- lected data and the stability of the learning process, this research area has attracted a lot of attention recently. In this chapter, we introduce the basic principles and the theory behind batch reinforcement learning, describe the most important algorithms, exemplarily discuss ongoing research within this field, and briefly survey real-world applications of batch reinforcement learning. 1 Introduction Batch reinforcement learning is a subfield of dynamic programming (DP) based re- inforcement learning (RL) that has vastly grown in importance during the last years. Historically, the term ‘batch RL’ is used to describe a reinforcement learning setting, where the complete amount of learning experience—usually a set of transitions sam- pled from the system—is fixed and given a priori (Ernst et al, 2005a). The task of the learning system then is to derive a solution—usually an optimal policy—out of this given batch of samples. In the following, we will relax this assumption of an a priori fixed set of training experience. The crucial benefit of batch algorithms lies in the way they handle a batch of transitions and get the best out of it, rather than in the fact that this set is fixed. From this perspective, batch RL algorithms are characterized by two basic constituents: all observed transitions are stored and updates occur synchronously on Sascha Lange, Thomas Gabel, Martin Riedmiller Albert-Ludwigs-Universtit ¨at Freiburg, Faculty of Engineering, Georges-K ¨ohler-Allee 079, D- 79110 Freiburg, Germany, e-mail: [slange,tgabel,riedmiller]@informatik.uni-freiburg.de 1

2210.15097.pdf

Contrastive Decoding: Open-ended Text Generation as Optimization Xiang Lisa Li1, Ari Holtzman2, Daniel Fried3, Percy Liang1, Jason Eisner4, Tatsunori Hashimoto1, Luke Zettlemoyer2,5, Mike Lewis5 Stanford University1, University of Washington2, Carnegie Mellon University3, Johns Hopkins University4, FAIR5 xlisali@stanford.edu ,ahai@cs.washington.edu ,dfried@cs.cmu.edu , pliang@stanford.edu ,jason@cs.jhu.edu ,thashim@stanford.edu , lsz@cs.washington.edu ,mikelewis@meta.com Abstract Given a language model (LM), maximum probability is a poor decoding objective for open-ended generation, because it produces short and repetitive text. On the other hand, sampling can often produce incoherent text that drifts from the original topics. We propose contrastive decoding (CD), a reliable decoding approach that optimizes a contrastive objective subject to a plausibility constraint. The contrastive objective returns the difference between the likelihood under a large LM (called the expert, e.g. OPT-13B) and a small LM (called the amateur, e.g. OPT-125M), and the constraint ensures that the outputs are plausible. CD is inspired by the fact that the failures of larger LMs (e.g., repetition, inco- herence) are even more prevalent in smaller LMs, and that this difference signals which texts should be preferred. CD requires zero additional training, and produces higher quality text than decoding from the larger LM alone. It also works across model scales (OPT-13B and GPT2-1.5B) and significantly outperforms four strong decoding algorithms (e.g., nucleus, top-k) in automatic and human evaluations across wikipedia, news and story domains.1 1 Introduction Open-ended text generation aims to craft fluent and coherent textual continuations of given prompts, laying foundations for various downstream applic- ations such as writing assistance and story gen- eration (Brown et al., 2020). The canonical ap- proaches often sample from large pre-trained lan- guage models (Holtzman et al., 2020; Fan et al., 2018; Radford et al., 2019), but the generated text is prone to incoherence and topic drift as unlucky sampling choices compound over long sequences (Eikema and Aziz, 2020; Maynez et al., 2020). On the other hand, searching for the most likely se- 1Code is available at https://github.com/ XiangLi1999/ContrastiveDecoding.git Figure 1: Contrastive decoding exploits the contrasts between expert and amateur LM of different sizes by choosing tokens that maximize their log-likelihood difference. CD produces high-quality text that amplifies the good expert behavior and diminishes the undesired amateur behavior. quences often results in short, repetitive and tedi- ous text (Holtzman et al., 2020), indicating that maximizing probability is a wrong decoding ob- jective. We propose a new search-based approach, contrastive decoding (CD), that can generate fluent and lexically diverse text without compromising coherence. As shown in Figure 1, contrastive decoding takes an off-the-shelf large language model such as OPT-13B (that we call the expert) and an off-the-shelf smaller language model such as OPT-125M (that we call the amateur). CD searches for text that maximizes the difference between expert log-probabilities and amateur log-probabilities, subject to plausibility constraints which restrict the search space to tokens with sufficiently high probability under the expert LM. Contrastive Decoding works because many fail- ure modes of language models (short, repetitive, ir- relevant or uninteresting strings) are more commonarXiv:2210.15097v2 [cs.CL] 10 Jul 2023

3639-the-effects-of-reward-misspeci.pdf

THEEFFECTS OF REWARD MISSPECIFICATION : MAPPING AND MITIGATING MISALIGNED MODELS Alexander Pan CaltechKush Bhatia UC BerkeleyJacob Steinhardt UC Berkeley ABSTRACT Reward hacking—where RL agents exploit gaps in misspecified reward functions—has been widely observed, but not yet systematically studied. To un- derstand how reward hacking arises, we construct four RL environments with misspecified rewards. We investigate reward hacking as a function of agent ca- pabilities: model capacity, action space resolution, observation space noise, and training time. More capable agents often exploit reward misspecifications, achiev- ing higher proxy reward and lower true reward than less capable agents. Moreover, we find instances of phase transitions : capability thresholds at which the agent’s behavior qualitatively shifts, leading to a sharp decrease in the true reward. Such phase transitions pose challenges to monitoring the safety of ML systems. To ad- dress this, we propose an anomaly detection task for aberrant policies and offer several baseline detectors. 1 I NTRODUCTION As reinforcement learning agents are trained with better algorithms, more data, and larger policy models, they are at increased risk of overfitting their objectives (Russell, 2019). Reward hacking , or the gaming of misspecified reward functions by RL agents, has appeared in a variety of con- texts, such as game playing (Ibarz et al., 2018), text summarization (Paulus et al., 2018), and au- tonomous driving (Knox et al., 2021). These examples show that better algorithms and models are not enough; for human-centered applications such as healthcare (Yu et al., 2019), economics (Trott et al., 2021) and robotics (Kober et al., 2013), RL algorithms must be safe and aligned with human objectives (Bommasani et al., 2021; Hubinger et al., 2019). Reward misspecifications occur because real-world tasks have numerous, often conflicting desider- ata. In practice, reward designers resort to optimizing a proxy reward that is either more readily measured or more easily optimized than the true reward. For example, consider a recommender system optimizing for users’ subjective well-being (SWB). Because SWB is difficult to measure, engineers rely on more tangible metrics such as click-through rates or watch-time. Optimizing for misspecified proxies led YouTube to overemphasize watch-time and harm user satisfaction (Stray, 2020), as well as to recommended extreme political content to users (Ribeiro et al., 2020). Addressing reward hacking is a first step towards developing human-aligned RL agents and one goal of ML safety (Hendrycks et al., 2021a). However, there has been little systematic work investigating when or how it tends to occur, or how to detect it before it runs awry. To remedy this, we study the problem of reward hacking across four diverse environments: traffic control (Wu et al., 2021), COVID response (Kompella et al., 2020), blood glucose monitoring (Fox et al., 2020), and the Atari game Riverraid (Brockman et al., 2016). Within these environments, we construct nine misspecified proxy reward functions (Section 3). Using our environments, we study how increasing optimization power affects reward hacking, by training RL agents with varying resources such as model size, training time, action space resolution, and observation space noise (Section 4). We find that more powerful agents often attain higher proxy reward but lower true reward, as illustrated in Figure 1. Since the trend in ML is to increase resources exponentially each year (Littman et al., 2021), this suggests that reward hacking will become more pronounced in the future in the absence of countermeasures. 1

2401.12187.pdf

WARM: On the Benefits of Weight Averaged Reward Models Alexandre Ramé, Nino Vieillard, Léonard Hussenot, Robert Dadashi, Geoffrey Cideron, Olivier Bachem, Johan Ferret Google DeepMind Aligning large language models (LLMs) with human preferences through reinforcement learning (RLHF) can lead to reward hacking, where LLMs exploit failures in the reward model (RM) to achieve seemingly high rewards without meeting the underlying objectives. We identify two primary challenges when designing RMs to mitigate reward hacking: distribution shifts during the RL process and inconsistencies in human preferences. As a solution, we propose Weight Averaged Reward Models ( WARM), first fine- tuning multiple RMs, then averaging them in the weight space. This strategy follows the observation that fine-tuned weights remain linearly mode connected when sharing the same pre-training. By averaging weights, WARMimproves efficiency compared to the traditional ensembling of predictions, while improving reliability under distribution shifts and robustness to preference inconsistencies. Our experiments on summarization tasks, using best-of- 𝑁and RL methods, shows that WARMimproves the overall quality and alignment of LLM predictions; for example, a policy RL fine-tuned with WARMhas a 79.4% win rate against a policy RL fine-tuned with a single RM. Keywords: Alignment, RLHF, Reward Modeling, Model Merging 1. Introduction Reward modeling. Conversational assistants such as Gemini [ 1] or GPT-4 [ 2] have revolutionized the AI community and beyond. These LLMs are capable of completing novel and intricate tasks, including mathematics, coding, and tool use [3]. These advancements are underpinned by a systematic three stage training procedure: pre-training by next token prediction [ 4,5,6], supervised fine-tuning (SFT) to learn to follow instructions [ 7,8,9], and ultimately, reinforcement learning (RL) to maximize a reward encapsulating the desired behaviors [ 10]. However, defining such rewards for real-world tasks is non-trivial [ 11]. In reinforcement learning from human feedback (RLHF) [ 12,13,14,15], rewards are reward models (RMs), trained on binary preference datasets to emulate human judgment. The enhancement of LLM capabilities from RL is strongly tied to the quality of the RMs [16]. Reward hacking. Particularly insidious in RLHF [ 17,18] is thereward hacking issue [19,20,21,22] (a.k.a. reward overoptimization), arising from reward misspecification [23,24] between the proxy RM and actual human preferences. While optimizing for the RM initially provides improvements, in later stages the policy (i.e., the LLM being trained) usually learns to exploit loopholes in the RM and achieves high rewards without truly fulfilling the intended objectives, as illustrated in Figure 1(b). Thisrewardhackingphenomenonposesnumerousissues. First, itdegradesperformances, manifesting as linguistically flawed [ 25] or unnecessarily verbose [ 26] outputs, which do not reflect true human preferences. Second, it complicates checkpoint selection due to the unreliability of the proxy RM, echoing Goodhart’s Law [ 27]: “when a measure becomes a target, it ceases to be a good measure”. Third, it can engender sycophancy [ 28,29] or amplify social biases, reflecting the limited and skewed demographics of feedback providers [ 30,31]. Lastly and most critically, misalignment [ 32,33] due to reward hacking can escalate into safety risks [ 19,34,35], in particular given the rapid integration of LLMs in everyday life and critical decision-making. Such concerns underscore the need to mitigate reward hacking to ensure the beneficial and safe deployment of LLMs. Corresponding author: alexandrerame@google.comarXiv:2401.12187v1 [cs.LG] 22 Jan 2024

2305.16183.pdf

Passive learning of active causal strategies in agents and language models Andrew K. Lampinen Google DeepMind London, UK lampinen@deepmind.comStephanie C. Y. Chan Google DeepMind London, UK scychan@deepmind.comIshita Dasgupta Google DeepMind London, UK idg@deepmind.com Andrew J. Nam Stanford University Stanford, CA ajhnam@stanford.eduJane X. Wang Google DeepMind London, UK wangjane@deepmind.com Abstract What can be learned about causality and experimentation from passive data? This question is salient given recent successes of passively-trained language models in interactive domains such as tool use. Passive learning is inherently limited. However, we show that purely passive learning can in fact allow an agent to learn generalizable strategies for determining and using causal structures, as long as the agent can intervene at test time. We formally illustrate that learning a strategy of first experimenting, then seeking goals, can allow generalization from passive learning in principle. We then show empirically that agents trained via imitation on expert data can indeed generalize at test time to infer and use causal links which are never present in the training data; these agents can also generalize experimentation strategies to novel variable sets never observed in training. We then show that strategies for causal intervention and exploitation can be generalized from passive data even in a more complex environment with high-dimensional observations, with the support of natural language explanations. Explanations can even allow passive learners to generalize out-of-distribution from perfectly- confounded training data. Finally, we show that language models, trained only on passive next-word prediction, can generalize causal intervention strategies from a few-shot prompt containing examples of experimentation, together with explanations and reasoning. These results highlight the surprising power of passive learning of active causal strategies, and may help to understand the behaviors and capabilities of language models. 1 Introduction Learning from passive observational data only allows learning correlational, not causal, structure. This observation is sometimes cited as a fundamental limitation of current machine learning research [52,53,34]. However, reinforcement learning (RL) agents can intervene on their environment, and are therefore not entirely limited. Indeed, various works have shown that RL agents can (meta-)learn to intervene on the environment to discover and exploit its causal structure [43, 13, 36, 15, 25]. However, these prior works leave open the possibility that an agent could passively learn a generaliz- able strategy for discovering and exploiting causal structure. While it is certainly necessary for an agent to intervene on the world at test time to discover causal structure, it may be possible for the agent to learn such a strategy from purely passive, offline data . Metaphorically, we ask “could an Preprint. Under review.arXiv:2305.16183v1 [cs.LG] 25 May 2023

2001.08361.pdf

Scaling Laws for Neural Language Models Jared Kaplan∗ Johns Hopkins University, OpenAI jaredk@jhu.eduSam McCandlish∗ OpenAI sam@openai.com Tom Henighan OpenAI henighan@openai.comTom B. Brown OpenAI tom@openai.comBenjamin Chess OpenAI bchess@openai.comRewon Child OpenAI rewon@openai.com Scott Gray OpenAI scott@openai.comAlec Radford OpenAI alec@openai.comJeffrey Wu OpenAI jeffwu@openai.comDario Amodei OpenAI damodei@openai.com Abstract We study empirical scaling laws for language model performance on the cross-entropy loss. The loss scales as a power-law with model size, dataset size, and the amount of compute used for training, with some trends spanning more than seven orders of magnitude. Other architectural details such as network width or depth have minimal effects within a wide range. Simple equations govern the dependence of overfitting on model/dataset size and the dependence of training speed on model size. These relationships allow us to determine the optimal allocation of a fixed compute budget. Larger models are significantly more sample- efficient, such that optimally compute-efficient training involves training very large models on a relatively modest amount of data and stopping significantly before convergence. ∗Equal contribution. Contributions: Jared Kaplan and Sam McCandlish led the research. Tom Henighan contributed the LSTM ex- periments. Tom Brown, Rewon Child, and Scott Gray, and Alec Radford developed the optimized Transformer implementation. Jeff Wu, Benjamin Chess, and Alec Radford developed the text datasets. Dario Amodei provided guidance throughout the project.arXiv:2001.08361v1 [cs.LG] 23 Jan 2020

10.1038.s41467-023-38539-w.pdf

Article https://doi.org/10.1038/s41467-023-38539-w A method for restoring signals and revealing individual macromolecule states incryo-ET, REST Haonan Zhang1,2,3,Y a nL i1,3,Y a n a nL i u1,2, Dongyu Li1,2,L i nW a n g1, Kai Song1, Keyan Bao1& Ping Zhu1,2 Cryo-electron tomography (cryo-ET) is widely used to explore the 3D density of biomacromolecules. However, the heavy noise and missing wedge effectprevent directly visualizing and analyzing the 3D reconstructions. Here, weintroduced REST, a deep learning stra tegy-based method to establish the relationship between low-quality and high-quality density and transfer theknowledge to restore signals in cryo-ET. Test results on the simulated and realcryo-ET datasets show that REST perform sw e l li nd e n o i s i n ga n dc o m p e n s a t i n g the missing wedge information. The application in dynamic nucleosomes,presenting either in the form of individ ual particles or in the context of cryo- FIB nuclei section, indicates that REST has the capability to reveal different conformations of target macromolecu les without subtomogram averaging. Moreover, REST noticeably improves the reliability of particle picking. Theseadvantages enable REST to be a powerfu l tool for the strai ghtforward inter- pretation of target macromolecules by vi sual inspection of the density and of a broad range of other applications in cryo-ET, such as segmentation, particlepicking, and subto mogram averaging. Cryo-ET has emerged as a powerful method which could record the 3D information of the biological macromolecules; however, many chal-lenges still remain to be addressed 1,2. First, the noise level of the tomogram is very high due to the radiation sensitivity of the samples, hence the low-dose electron tomography hinders human eyes toidentify the features in it 3. Second, during the data collection, tilt-series images can only be collected within a tilt angular range of approxi-mately ±70° because of the limitation of the specimen holder. Thiscould lead to incomplete 3D information in the Fourier space, resultingin a so-called missing wedge in the tomogram. The effect of themissing wedge is clearly visible in the 3D Fourier transform of the beamdirection. The most obvious artefact caused by a missing wedge is theanisotropic resolution, in which objects appear elongated in thedirection of the beam axis, i.e., in the Z direction 4. The EM density in the 3D and 2D slices related to the Z-plane are distorted as a result ofthis elongation. Therefore, most of 3D segmentation was unable to entail in Z direction and render a highlight extended structure. To address these challenges in cryo-ET, a variety of methods have been proposed to recover the information and produce high contrast tomograms5. During the data collection, dual-axis tomography, in which the tilt series are collected using two perpendicular axes, couldbe applied 6. However, this method is limited by the use of a higher electron dose, which may damage the biological specimen7.I no t h e r studies that have focused on the data processing procedures, a seriesof algorithms, including the algebraic reconstruction technique(ART) 8,s i m u l t a n e o u sA R T( S A R T )9and simultaneous iterative recon- struction technique (SIRT)10, have been proposed to improve the quality of tomograms. These methods, which are mainly based onmathematic calculations, reduce the differences between the calcu-lated projections of the reconstructed tomogram and the tilt series. ByReceived: 4 August 2022 Accepted: 8 May 2023 Check for updates 1National Laboratory of Biomacromolecules, CAS Center for Excellence in Bio macromolecules, Institute of Biophysics, Chinese Academy of Sciences, Beijing 100101, China.2University of Chinese Academy of Sciences, Beijing 100049, China.3These authors contributed equally: Haonan Zhang, Yan Li. e-mail: zhup@ibp.ac.cn Nature Communications | (2023) 14:2937 11234567890():,; 1234567890():,;

1801.10198.pdf

Published as a conference paper at ICLR 2018 GENERATING WIKIPEDIA BY SUMMARIZING LONG SEQUENCES Peter J. Liu∗, Mohammad Saleh∗, Etienne Pot†, Ben Goodrich, Ryan Sepassi, Łukasz Kaiser, Noam Shazeer Google Brain Mountain View, CA {peterjliu,msaleh,epot,bgoodrich,rsepassi,lukaszkaiser,noam }@google.com ABSTRACT We show that generating English Wikipedia articles can be approached as a multi- document summarization of source documents. We use extractive summarization to coarsely identify salient information and a neural abstractive model to generate the article. For the abstractive model, we introduce a decoder-only architecture that can scalably attend to very long sequences, much longer than typical encoder- decoder architectures used in sequence transduction. We show that this model can generate fluent, coherent multi-sentence paragraphs and even whole Wikipedia articles. When given reference documents, we show it can extract relevant factual information as reflected in perplexity, ROUGE scores and human evaluations. 1 I NTRODUCTION The sequence-to-sequence framework has demonstrated success in natural-language sequence trans- duction tasks such as machine translation. More recently, neural techniques have been applied to do single-document, abstractive (paraphrasing) text summarization of news articles (Rush et al. (2015), Nallapati et al. (2016)). In this prior work, the input to supervised models ranged from the first sen- tence to the entire text of an article, and they are trained end-to-end to predict reference summaries. Doing this end-to-end requires a significant number of parallel article-summary pairs since language understanding is a pre-requisite to generate fluent summaries. In contrast, we consider the task of multi-document summarization, where the input is a collection of related documents from which a summary is distilled. Prior work has focused on extractive summarization, which select sentences or phrases from the input to form the summaries, rather than generating new text. There has been limited application of abstractive neural methods and one possible reason is the paucity of large, labeled datasets. In this work, we consider English Wikipedia as a supervised machine learning task for multi- document summarization where the input is comprised of a Wikipedia topic (title of article) and a collection of non-Wikipedia reference documents, and the target is the Wikipedia article text. We describe the first attempt to abstractively generate the first section, or lead, of Wikipedia articles con- ditioned on reference text. In addition to running strong baseline models on the task, we modify the Transformer architecture (Vaswani et al., 2017) to only consist of a decoder, which performs better in the case of longer input sequences compared to recurrent neural network (RNN) and Transformer encoder-decoder models. Finally we show our modeling improvements allow us to generate entire Wikipedia articles. ∗Joint first-authors. Ordered randomly. †Work done as a member of the Google Brain Residency (g.co/brainresidency) 1arXiv:1801.10198v1 [cs.CL] 30 Jan 2018

10.1126.science.abo7201.pdf

RESEARCH ARTICLE SUMMARY◥ CORONAVIRUS Open science discovery of potent noncovalent SARS-CoV-2 main protease inhibitors Melissa L. Boby †, Daren Fearon †, Matteo Ferla †, Mihajlo Filep †, Lizbé Koekemoer †, Matthew C. Robinson †, The COVID Moonshot Consortium, John D. Chodera *, Alpha A. Lee *, Nir London *, Annette von Delft *, Frank von Delft * INTRODUCTION: COVID-19 became a global pan- demic partially as a result of the lack of easily deployable, broad-spectrum oral antivirals, which complicated its containment. Even en-demically, and with effect ive vaccinations, it will continue to cause acute disease, death, and long- term sequelae globally unless there are acces- sible treatments. COVID-19 is not an isolated event but instead is the latest example of a viral pandemic threat to human health. Therefore, antiviral discovery and development should be a key pillar of pandemic preparedness efforts. RATIONALE: One route to accelerate antiviral drug discovery is the establishment of open knowledge bases, the development of effective technology infrastructures, and the discovery of multiple potent antivi rals suitable as start- ing points for the development of therapeu- tics. In this work, we report the results of the COVID Moonshot —a fully open science, crowd- sourced, and structure-enabled drug discovery campaign —against the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) main protease (Mpro). This collaboration may serve as a roadmap for the potential development of future antivirals. RESULTS: On the basis of the results of a crys- tallographic fragment screen, we crowdsourceddesign ideas to progress from fragment to lead compounds. The crowdsourcing strat- egy yielded several key compounds along the optimization trajectory, including the startingcompound of what became the primary lead series. Three additional chemically distinct lead series were also explored, spanning a di- versity of chemotypes. T h ec o l l a b o r a t i v ea n dh i g h l ya u t o m a t e dn a t u r e of the COVID Moonshot Consortium resulted in >18,000 compound designs, >2400 synthesized compounds, >490 ligand-bound x-ray structures, >22,000 alchemical free-energy calculations, and >10,000 biochemical measurements —all of which were made publicly available in real time. The recently approved antiviral ensitrelvir was identified in part bas ed on crystallographic data from the COVID Moonshot Consortium. This campaign led to the discovery of a po- tent [median inhibitory concentration (IC 50)= 37 ± 2 nM] and differentiated (noncovalent and nonpeptidic) lead compou nd that also exhibited potent cellular activity, with a median effective concentration (EC 50) of 64 nM in A549-ACE2- TMPRSS2 cells and 126 nM in HeLa-ACE2 cells without measurable cytotoxicity. Although the pharmacokinetics of th er e p o r t e dc o m p o u n di s not yet optimal for therapeutic development, it is a promising starting point for further antiviral discovery and development.CONCLUSION: T h es u c c e s so ft h eC O V I DM o o n - shot project in producing potent antivirals, building open knowledge bases, accelerating ex- ternal discovery efforts, and functioning as a useful information-exchange hub is an example of the potential effectiveness of open science antiviral discovery programs. The open science, patent-free nature of the project enabled a large number of collaborators to provide in-kind sup- port, including synthesis, assays, and in vitro and in vivo experiments. By making all data imme- diately available and ensuring that all compounds are purchasable from Enamine without the need for materials transfer agreements, we aim to ac-celerate research globally along parallel tracks. In the process, we generated a detailed map of the structural plasticity of Mpro, extensive structure-activity relationships for multiple chemotypes, and a wealth of biochemical activ- ity data to spur further research into antivirals and discovery methodologies. We hope that this can serve as an alternative model for antiviral discovery and future pandemic preparedness. Further, the project also showcases the role of machine learning, computational chemistry, and high-throughput structural biology as force mul- tipliers in drug design. Art ificial intelligence and machine learning algorithms help accelerate chemical synthesis while balancing multiple com- peting molecular properties. The design-make-test- analyze cycle was accelerated by these algorithms combined with planetary-scale biomolecular sim- ulations of protein-ligand interactions and rapid structure determination.▪RESEARCH The list of author affiliations is available in the full article online. *Corresponding author. Email: john.chodera@choderalab.org(J.D.C.); alpha.lee@postera.ai (A.A.L.); nir.london@weizmann. ac.il (N.L.); annette.vondelft@cmd.ox.ac.uk (A.v.D.); frank.von-delft@diamond.ac.uk (F.v.D.) †These authors contributed equally to this work. Cite this article as M. L. Boby et al .,Science 382, eabo7201 (2023). DOI: 10.1126/science.abo7201 READ THE FULL ARTICLE AT https://doi.org/10.1126/science.abo7201 CrowdsourcingMulti-institute collaboration Accelerated design-make-test cyclesRoute prediction Alchemicalfree-energycalculations >18,000 designs>2400 synthesizedHigh-throughput crystallography High-throughput assays >490 structures>10,000 measurements+Open dataMAT-POS-e194df51-1 Oralhalf-life: 1.4 h37 nM 64 nM NNH O S OO CN ClNCOVID Moonshot The COVID Moonshot Consortium. An open science, crowdsourced drug discovery campaign against the SARS-CoV-2 Mpro led to a potent, noncovalent, and nonpeptidic inhibitor scaffold with lead-like properties. We generated copious structural, biochemical, and pharmacological data that were shar ed rapidly and openly, creating a rich, open, and intellectual property –free knowledge base for future anticoronavirus drug discovery.CREDIT: ICONS MADE BY FREEPIK AND GOOD WARE FROM WWW.FLATICON.COM Boby et al.,Science 382, 663 (2023) 10 November 2023 1o f1 Downloaded from https://www.science.org on November 18, 2023

2306.16410.pdf

Towards Language Models That Can See: Computer Vision Through the LENS of Natural Language William Berrios†Gautam Mittal†§Tristan Thrush†§ Douwe Kiela†§Amanpreet Singh† †Contextual AI;§Stanford University Abstract We propose LENS , a modular approach for tackling computer vision problems by leveraging the power of large language models (LLMs). Our system uses a language model to reason over outputs from a set of independent and highly descriptive vision modules that provide exhaustive information about an image. We evaluate the approach on pure computer vision settings such as zero- and few-shot object recognition, as well as on vision and language problems. LENS can be applied to any off-the-shelf LLM and we find that the LLMs with LENS perform highly competitively with much bigger and much more sophisticated systems, without any multimodal training whatsoever. We open-source our code at https://github.com/ContextualAI/lens and provide an interactive demo1. Pretrained and frozen Trained from scratch (a) Multimodal Pretraining (b) No Multimodal PretrainingArchitectureData Source LENS (Ours) No additional pre-training data Flamingo Frozen LLM LM BlockXATTN LayerXATTN LayerSurfing Q: What is the dog doing?Image EncoderOutput Text LM Block Perceiver M3W 43M webpages ~2B samples image-video text pairs BLIP-2Q- FormerFC LayerSurfing Q: What is the dog doing?Image Encoder Frozen LLMOutput Text COCO Visual Genome CC12M SBU LAION-400M 115M images + synthetic captions Old-Style PretrainingSurfingOutput Text Q: What is the dog doing?Image Encoder Text Encoder Cross - Modality Encoder COCO Visual Genome VQA GQA Visual7W ... Millions of paired image/text samples Visual DescriptorsAttributesObjects CaptionsFrozen LLM Q: What is the dog doing?SurfingOutput Text Figure 1: Comparison of approaches for aligning visual and language modalities: (a) Multimodal pretraining using a paired or web dataset, and (b) LENS , a pretraining-free method that can be applied to any off-the-shelf LLM without the need for additional multimodal datasets. Unlike LENS, prior methods are computationally intensive and require joint alignment pretraining on large multimodal datasets to perform visual tasks. 1https://lens.contextual.ai/ Correspondence to lens@contextual.ai.arXiv:2306.16410v1 [cs.CL] 28 Jun 2023

2005.00341.pdf

Jukebox: A Generative Model for Music Prafulla Dhariwal* 1Heewoo Jun* 1Christine Payne* 1Jong Wook Kim1Alec Radford1Ilya Sutskever1 Abstract We introduce Jukebox, a model that generates music with singing in the raw audio domain. We tackle the long context of raw audio using a multi- scale VQ-V AE to compress it to discrete codes, and modeling those using autoregressive Trans- formers. We show that the combined model at scale can generate high-fidelity and diverse songs with coherence up to multiple minutes. We can condition on artist and genre to steer the musical and vocal style, and on unaligned lyrics to make the singing more controllable. We are releasing thousands of non cherry-picked samples, along with model weights and code. 1. Introduction Music is an integral part of human culture, existing from the earliest periods of human civilization and evolving into a wide diversity of forms. It evokes a unique human spirit in its creation, and the question of whether computers can ever capture this creative process has fascinated computer scien- tists for decades. We have had algorithms generating piano sheet music (Hiller Jr & Isaacson, 1957; Moorer, 1972; Hadjeres et al., 2017; Huang et al., 2017), digital vocoders generating a singer’s voice (Bonada & Serra, 2007; Saino et al., 2006; Blaauw & Bonada, 2017) and also synthesizers producing timbres for various musical instruments (Engel et al., 2017; 2019). Each captures a specific aspect of music generation: melody, composition, timbre, and the human voice singing. However, a single system to do it all remains elusive. The field of generative models has made tremendous progress in the last few years. One of the aims of gen- erative modeling is to capture the salient aspects of the data and to generate new instances indistinguishable from the true data The hypothesis is that by learning to produce the data we can learn the best features of the data1. We are surrounded by highly complex distributions in the visual, audio, and text domain, and in recent years we have devel- *Equal contribution1OpenAI, San Francisco. Correspondence to: <jukebox@openai.com>.oped advances in text generation (Radford et al.), speech generation (Xie et al., 2017) and image generation (Brock et al., 2019; Razavi et al., 2019). The rate of progress in this field has been rapid, where only a few years ago we had algorithms producing blurry faces (Kingma & Welling, 2014; Goodfellow et al., 2014) but now we now can gener- ate high-resolution faces indistinguishable from real ones (Zhang et al., 2019b). Generative models have been applied to the music genera- tion task too. Earlier models generated music symbolically in the form of a pianoroll, which specifies the timing, pitch, velocity, and instrument of each note to be played. (Yang et al., 2017; Dong et al., 2018; Huang et al., 2019a; Payne, 2019; Roberts et al., 2018; Wu et al., 2019). The symbolic approach makes the modeling problem easier by working on the problem in the lower-dimensional space. However, it constrains the music that can be generated to being a specific sequence of notes and a fixed set of instruments to render with. In parallel, researchers have been pursuing the non- symbolic approach, where they try to produce music directly as a piece of audio. This makes the problem more challeng- ing, as the space of raw audio is extremely high dimensional with a high amount of information content to model. There has been some success, with models producing piano pieces either in the raw audio domain (Oord et al., 2016; Mehri et al., 2017; Yamamoto et al., 2020) or in the spectrogram domain (Vasquez & Lewis, 2019). The key bottleneck is that modeling the raw audio directly introduces extremely long-range dependencies, making it computationally chal- lenging to learn the high-level semantics of music. A way to reduce the difficulty is to learn a lower-dimensional encod- ing of the audio with the goal of losing the less important information but retaining most of the musical information. This approach has demonstrated some success in generat- ing short instrumental pieces restricted to a set of a few instruments (Oord et al., 2017; Dieleman et al., 2018). In this work, we show that we can use state-of-the-art deep generative models to produce a single system capable of gen- erating diverse high-fidelity music in the raw audio domain, with long-range coherence spanning multiple minutes. Our approach uses a hierarchical VQ-V AE architecture (Razavi 1Richard Feynmann famously said, “What I cannot create, I do not understand”

1905.01969v4.pdf

Published as a conference paper at ICLR 2020 Poly-encoders :architectures and pre -training strategies for fast and accurate multi -sentence scoring Samuel Humeau∗, Kurt Shuster∗, Marie-Anne Lachaux, Jason Weston Facebook AI Research {samuelhumeau,kshuster,malachaux,jase }@fb.com Abstract The use of deep pre-trained transformers has led to remarkable progress in a num- ber of applications (Devlin et al., 2019). For tasks that make pairwise compar- isons between sequences, matching a given input with a corresponding label, two approaches are common: Cross-encoders performing full self-attention over the pair and Bi-encoders encoding the pair separately. The former often performs better, but is too slow for practical use. In this work, we develop a new trans- former architecture, the Poly-encoder , that learns global rather than token level self-attention features. We perform a detailed comparison of all three approaches, including what pre-training and fine-tuning strategies work best. We show our models achieve state-of-the-art results on four tasks; that Poly-encoders are faster than Cross-encoders and more accurate than Bi-encoders; and that the best results are obtained by pre-training on large datasets similar to the downstream tasks. 1 I ntroduction Recently, substantial improvements to state-of-the-art benchmarks on a variety of language under- standing tasks have been achieved through the use of deep pre-trained language models followed by fine-tuning (Devlin et al., 2019). In this work we explore improvements to this approach for the class of tasks that require multi-sentence scoring: given an input context, score a set of candidate labels, a setup common in retrieval and dialogue tasks, amongst others. Performance in such tasks has to be measured via two axes: prediction quality and prediction speed, as scoring many candidates can be prohibitively slow. The current state-of-the-art focuses on using BERT models for pre-training (Devlin et al., 2019), which employ large text corpora on general subjects: Wikipedia and the Toronto Books Corpus (Zhu et al., 2015). Two classes of fine-tuned architecture are typically built on top: Bi-encoders and Cross-encoders. Cross-encoders (Wolf et al., 2019; Vig & Ramea, 2019), which perform full (cross) self-attention over a given input and label candidate, tend to attain much higher accuracies than their counterparts, Bi-encoders (Mazar ´e et al., 2018; Dinan et al., 2019), which perform self-attention over the input and candidate label separately and combine them at the end for a final representa- tion. As the representations are separate, Bi-encoders are able to cache the encoded candidates, and reuse these representations for each input resulting in fast prediction times. Cross-encoders must recompute the encoding for each input and label; as a result, they are prohibitively slow at test time. In this work, we provide novel contributions that improve both the quality and speed axes over the current state-of-the-art. We introduce the Poly-encoder, an architecture with an additional learnt at- tention mechanism that represents more global features from which to perform self-attention, result- ing in performance gains over Bi-encoders and large speed gains over Cross-Encoders. To pre-train our architectures, we show that choosing abundant data more similar to our downstream task also brings significant gains over BERT pre-training. This is true across all di fferent architecture choices and downstream tasks we try. We conduct experiments comparing the new approaches, in addition to analysis of what works best for various setups of existing methods, on four existing datasets in the domains of dialogue and in- formation retrieval (IR), with pre-training strategies based on Reddit (Mazar ´e et al., 2018) compared ∗Joint First Authors. 1arXiv:1905.01969v4 [cs.CL] 25 Mar 2020

2401.18079.pdf

KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization Coleman Hooper chooper@berkeley.edu UC BerkeleySehoon Kim sehoonkim@berkeley.edu UC BerkeleyHiva Mohammadzadeh hiva@berkeley.edu UC Berkeley Michael W. Mahoney mmahoney@stat.berkeley.edu ICSI, LBNL, UC BerkeleyYakun Sophia Shao ysshao@berkeley.edu UC BerkeleyKurt Keutzer keutzer@berkeley.edu UC Berkeley Amir Gholami amirgh@berkeley.edu ICSI, UC Berkeley ABSTRACT LLMs are seeing growing use for applications which require large context windows, and with these large context windows KV cache activations surface as the dominant contributor to memory con- sumption during inference. Quantization is a promising approach for compressing KV cache activations; however, existing solutions fail to represent activations accurately in sub-4-bit precision. Our work, KVQuant, facilitates low precision KV cache quantization by incorporating several novel methods: (i) Per-Channel Key Quan- tization , where we adjust the dimension along which we quan- tize the Key activations to better match the distribution; (ii) Pre- RoPE Key Quantization , where we quantize Key activations before the rotary positional embedding to mitigate its impact on quan- tization; (iii) Non-Uniform KV Cache Quantization , where we de- rive per-layer sensitivity-weighted non-uniform datatypes that better represent the distributions; (iv) Per-Vector Dense-and-Sparse Quantization , where we isolate outliers separately for each vec- tor to minimize skews in quantization ranges; and (v) Q-Norm , where we normalize quantization centroids in order to mitigate distribution shift, providing additional benefits for 2-bit quantiza- tion. By applying our method to the LLaMA, LLaMA-2, and Mis- tral models, we achieve <0.1perplexity degradation with 3-bit quantization on both Wikitext-2 and C4, outperforming existing approaches. Our method enables serving LLaMA-7B with a con- text length of up to 1 million on a single A100-80GB GPU and up to 10 million on an 8-GPU system . We develop cus- tom CUDA kernels for KVQuant, showing that we can achieve up to∼1.4×speedups, compared to baseline fp16 matrix-vector multiplications, for the LLaMA-7B model. The code is available at https://github.com/SqueezeAILab/KVQuant/. 1 INTRODUCTION Large language models (LLMs) have revolutionized many natural language processing (NLP) tasks. In order to improve the capabili- ties of LLMs, there is significant interest in increasing the context lengths of LLMs. Longer context lengths enable new applications, including long document summarization, retrieval for answering questions about long documents, extended multi-turn applications [4], and code analysis. To support this pull from applications, therehave been significant recent advances in long-context length mod- els in industry [1, 27], as well as in academia [4]. Given the importance of LLM workloads, there is strong motiva- tion to improve their inference efficiency. LLM inference with large context lengths can be incredibly resource-intensive; serving LLMs requires high-end GPUs, and the largest LLMs require costly multi- GPU inference setups. When analyzing the computational nature of generative inference with LLMs, it becomes quickly apparent that, for relatively small batch sizes, the computation is memory bound [18]. With the growing divergence between computational speeds and memory speeds, this problem is only going to get worse over time [ 13]. This makes reducing the memory bottleneck preem- inently important. Further analysis shows that the memory bottle- neck is strongly related to context size. For short sequence lengths, the dominant contributor to memory consumption is the weight ma- trices, and therefore the optimal strategy is to minimize the model size in order to reduce memory consumption as well as bandwidth requirements [ 18,19]. However, for long sequence lengths, the main bottleneck is the memory requirements for caching Key and Value (KV) activations throughout inference. In particular, the size of the KV cache can become the dominant contributor to memory footprint, even for a 32K context limit (see Table 1), making it chal- lenging to perform long context length inference. This challenge is further exacerbated when one considers batched inference. It is therefore crucial to develop methods for compressing the KV cache to enable efficient long-sequence length inference. Existing approaches lead to unacceptable accuracy degradation due to the outlier structures in KV cache activations as well as suboptimal bit allocation with existing uniform and non-uniform approaches. In this work, we perform an extensive analysis of KV cache activa- tions in recent LLMs, revealing patterns which can be exploited to enable ultra-low precision quantization with minimal accuracy loss. In particular, we make the following contributions (summarized in Figure 1): •We find that the Key matrices exhibit structured outliers in spe- cific channels before applying RoPE. However, the outlier channel magnitudes become less consistent after applying RoPE, posing a distinct challenge for low precision quantization. Based on these observations, we use per-channel quantization for Keys,arXiv:2401.18079v2 [cs.LG] 7 Feb 2024

2305.15717.pdf

The False Promise of Imitating Proprietary LLMs Arnav Gudibande∗ UC Berkeley arnavg@berkeley.eduEric Wallace∗ UC Berkeley ericwallace@berkeley.eduCharlie Snell∗ UC Berkeley csnell22@berkeley.edu Xinyang Geng UC Berkeley young.geng@berkeley.eduHao Liu UC Berkeley hao.liu@berkeley.eduPieter Abbeel UC Berkeley pabbeel@berkeley.edu Sergey Levine UC Berkeley svlevine@berkeley.eduDawn Song UC Berkeley dawnsong@berkeley.edu Abstract An emerging method to cheaply improve a weaker language model is to finetune it on outputs from a stronger model, such as a proprietary system like ChatGPT (e.g., Alpaca, Self-Instruct, and others). This approach looks to cheaply imitate the proprietary model’s capabilities using a weaker open-source model. In this work, we critically analyze this approach. We first finetune a series of LMs that imitate ChatGPT using varying base model sizes (1.5B–13B), data sources, and imitation data amounts (0.3M–150M tokens). We then evaluate the models using crowd raters and canonical NLP benchmarks. Initially, we were surprised by the output quality of our imitation models—they appear far better at following instructions, and crowd workers rate their outputs as competitive with ChatGPT. However, when conducting more targeted automatic evaluations, we find that imitation models close little to none of the gap from the base LM to ChatGPT on tasks that are not heavily supported in the imitation data. We show that these performance discrepancies may slip past human raters because imitation models are adept at mimicking ChatGPT’s style but not its factuality . Overall, we conclude that model imitation is a false promise: there exists a substantial capabilities gap between open and closed LMs that, with current methods, can only be bridged using an unwieldy amount of imitation data or by using more capable base LMs. In turn, we argue that the highest leverage action for improving open-source models is to tackle the difficult challenge of developing better base LMs, rather than taking the shortcut of imitating proprietary systems. 1 Introduction The recent release of powerful language models (LMs) such as ChatGPT (OpenAI, 2022), Bard (Pichai, 2023), and Claude (AnthropicAI, 2023) might herald a future where the best AI systems are provided primarily as a fee-based API by large companies. At the same time, open-source LMs are becoming increasingly accurate, with models like LLaMA and FLAN-T5 providing many of the same basic capabilities as their commercial counterparts, albeit at a lower level of perfor- mance (Touvron et al., 2023; Chung et al., 2022). This presents an important question, whose answer will have profound future implications: will the most powerful LMs be closed-source or will they be freely distributed for anyone to use, modify, and extend? Both possibilities have important pros and cons, and implications on policy, corporate strategy, and the future of scientific inquiry. ∗Equal Contribution. Preprint. Under review.arXiv:2305.15717v1 [cs.CL] 25 May 2023

2306.02707.pdf

Orca: Progressive Learning from Complex Explanation Traces of GPT-4 Subhabrata Mukherjee∗†, Arindam Mitra∗ Ganesh Jawahar, Sahaj Agarwal, Hamid Palangi, Ahmed Awadallah Microsoft Research Abstract Recent research has focused on enhancing the capability of smaller models through imitation learning, drawing on the outputs generated by large foundation models (LFMs). A number of issues impact the quality of these models, ranging from limited imitation signals from shallow LFM outputs; small scale homogeneous training data; and most notably a lack of rigorous evaluation resulting in overestimating the small model’s capability as they tend to learn to imitate the style, but not the reasoning process of LFMs . To address these challenges, we develop Orca, a 13-billion parameter model that learns to imitate the reasoning process of LFMs. Orca learns from rich signals from GPT-4 including explanation traces; step-by-step thought processes; and other complex instructions, guided by teacher assistance from ChatGPT. To promote this progressive learning, we tap into large-scale and diverse imitation data with judicious sampling and selection. Orca surpasses conventional state-of-the-art instruction-tuned models such as Vicuna-13B by more than 100% in complex zero-shot reasoning benchmarks like Big- Bench Hard (BBH) and 42%on AGIEval. Moreover, Orca reaches parity with ChatGPT on the BBH benchmark and shows competitive performance (4pts gap with optimized system message) in professional and academic examinations like the SAT, LSAT, GRE, and GMAT, both in zero-shot settings without CoT; while trailing behind GPT-4. Our research indicates that learning from step-by-step explanations, whether these are generated by humans or more advanced AI models, is a promising direction to improve model capabilities and skills. ∗Co-primary authors. Author contributions listed at the end of the paper. †Correspondence to subhabrata.mukherjee@microsoft.com We are working with our legal team to publicly release a diff of the model weights in accordance with LLaMA’s release policy to be published at https://aka.ms/orca-lm . Work in progress.arXiv:2306.02707v1 [cs.CL] 5 Jun 2023

109_how_well_do_generative_protein.pdf

HOW WELL DO GENERATIVE PROTEIN MODELS GENERATE ? Han Spinner Department of Systems Biology Harvard Medical SchoolAaron W. Kollasch Department of Systems Biology Harvard Medical SchoolDebora S. Marks Department of Systems Biology Harvard Medical School ABSTRACT Protein design relies critically on the generation of plausible sequences. Yet, the efficacy of many common model architectures from simple interpretable models, like position- specific scoring matrix (PSSM) and direct couplings analysis (DCA), to newer and less interpretable models, like variational autoencoders (V AEs), autoregressive large language models (AR-LLMs) and flow matching (FM), for sequence sampling remains uncertain. While some models offer unique sequence generation methods, issues such as mode col- lapse, generation of nonsensical repeats, and protein truncations persist. Trusted methods like Gibbs sampling are often preferred for their reliability, but can be computationally expensive. This paper addresses the need to evaluate the performance and limitations of different generation methods from protein models, considering dependencies on multiple sequence alignment (MSA) depth and available sequence diversity. We propose rigorous evaluation methods and metrics to assess sequence generation, aiming to guide design de- cisions and inform the development of future model and sampling techniques for protein design applications. 1 I NTRODUCTION Using machine learning to design proteins is useless unless we can generate plausible sequences, regardless of training data or model type. Many different approaches to protein design to achieve different goals have been quite successful (Shin et al., 2021; Madani et al., 2023; Lian et al., 2022; Hawkins-Hooker et al., 2021), and all of these projects have hinged on generating sequences that ‘make sense’. In almost all protein engineering and protein design quests, we want to create proteins that fold and function. However, conditions that encourage stability, dynamic movements, tolerance to stressors, proper expression levels, etc, are often specific protein-to-protein and project-to-project. In order to increase efficacy of these studies we must ask the simple question: How well do generative protein models generate? Newer model architectures, such as variational autoencoders (V AE) and autoregressive large language mod- els (AR-LLM), have shown some promise for function and structure prediction Frazer et al. (2021); Hsu et al. (2022); Notin et al. (2022). And the importance of comparing to simpler, more interpretable models has also been noted Zhang et al. (2024). Often, However, there are no theoretical guarantees that models successful for structure or fitness predictions are also guaranteed to be better for sampling new sequences. These models’ architectures enable unique ways of generating: for instance, sampling sequences from a learned latent space from a V AE and ancestral sampling for AR-LLMs. But often, there are malignancies that come from these generation methods that go ignored. When drawing sequences from the latent space of a V AE, it is common to observe issues with the diversity of sequences that are generated: (a) mode collapse, (b) posterior collapse, (c) even distribution of sequence diversity across the length of the protein, and/or (d) low quality sequences that have mutated active site or other key residues (Figure 1a). Examples of malignancies from ancestral sampling from AR-LLMs include: 1

Pursuing-structural-biology-in-China-cell.pdf

Leading Edge Conversations Pursuing structural biology in China In November 2023, structural biologists from different countries and different disciplines gathered at the Cell Symposium: Structural biology from the nanoscale to cellular mesoscale to discuss recent breakthroughs,including structures of proteins and macromolecular complexes in a cellular context as well as virus struc-tures obtained by using different techniques. At the symposium, Cell editor Jia Cheng and Karin Ku ¨ hnel, editor-in-chief of Structure , spoke with Drs. Beili Wu, Mingjie Zhang, and Zihe Rao about their experiences doing structural biology research in China and about their perspectives for the future. An edited transcriptof the conversation is presented below, and the full conversation is available with the article online. Jia Cheng: It’s my great pleasure to have Dr. Mingjie Zhang, Dr. Beili Wu [and later Dr. Zihe Rao] join this conversation. I would like to start with the first question. Could you tell us abouthow each of you got interested in structural biology or structural neuroscience?Beili Wu: I got interested in structural biology after I joined Professor Rao’s lab in Tsinghua University as a PhD student. I was impressed by the beauty of protein crystals. It’s like themost beautiful jewelry that I can grow by myself. Later, I was fascinated by the logic of structures because everything Figure 1(L to R) Jia Cheng, Karin Ku ¨ hnel, Zihe Rao, Mingjie Zhang, Beili Wu ll Cell 187, February 1, 2024 ª2024 Elsevier Inc. 513

HyvO00-icatut.pdf

Indep enden t Comp onen t Analysis/: A T utorialAap o Hyv /ärinen and Erkki OjaHelsinki Univ ersit y of T ec hnologyLab oratory of Computer and Information ScienceP /.O/. Bo x /5/4/0/0/, FIN/-/0/2/0/1/5 Esp o o/, Finlandaapo/.hyvarinen/@hut/.fi/, erkki/.oja/@hut/.fihttp/:////www/.cis/.hut/.fi//pro ject s//ic a//A v ersion of this pap er will app ear in Neur al Networkswith the title /Indep enden t Comp onen tA n a l ysis/: Algorithms and Applications/April /1/9/9/9/1 Motiv ationImagine that y ou are in a ro om where t w op eople are sp eaking sim ultaneously /. Y ou ha v et w om icrophones/,whic hy ou hold in di/eren tl o c a t ions/. The microphones giv ey ou t w o recorded time signals/, whic hw e coulddenote b y x/1 /( t /) and x/2 /( t /) /, with x/1 and x/2 the amplitudes/, and t the time index/. Eac ho f these recordedsignals is a w eigh ted sum of the sp eec hs ignals emitted b yt h et w os p eak ers/, whic hw ed enote b y s/1 /( t /) ands/2 /( t /) /. W e could express this as a linear equation/:x/1 /( t /)/= a/1/1 s/1 /+ a/1/2 s/2 /(/1/)x/2 /( t /)/= a/2/1 s/1 /+ a/2/2 s/2 /(/2/)where a/1/1 /;;a/1/2 /;;a/2/1 /, and a/2/2 are some parameters that dep end on the distances of the microphones fromthe sp eak ers/. It w ould b e v ery useful if y ou could no w estimate the t w oo r i ginal sp eec h signals s/1 /( t /) ands/2 /( t /) /, using only the recorded signals x/1 /( t /) and x/2 /( t /) /. This is called the c o cktail/-p arty pr oblem /. F or thetime b eing/, w eo m i ta n y time dela ys or other extra factors from our simpli/ed mixing mo del/.As an illustration/, consider the w a v eforms in Fig/. /1 and Fig/. /2/. These are/, of course/, not realistic sp eec hsignals/, but su/ce for this illustration/. The original sp eec hs ignals could lo ok something lik et hose in Fig/. /1and the mixed signals could lo ok lik et h o s ei n Fig/. /2/. The problem is to reco v er the data in Fig/. /1 usingonly the data in Fig/. /2/.A ctually /,i f w e knew the parameters aij /,w e could solv et he linear equation in /(/1/) b y classical metho ds/.The p oin ti s /, ho w ev er/, that if y ou don/'t kno wt h e aij /, the problem is considerably more di/cult/.One approac ht o solving this problem w ould b e to use some information on the statistical prop erties ofthe signals si /( t /) to estimate the aii /. A ctually /, and p erhaps surprisingly /,i tt urns out that it is enough toassume that s/1 /( t /) and s/2 /( t /) /,a t e a c ht i m e i nstan t t /,a r e statistic al ly indep endent /. This is not an unrealisticassumption in man yc ases/, and it need not b e exactly true in practice/. The recen tly dev elop ed tec hniqueof Indep enden tC omp onen tA nalysis/, or ICA/, can b e used to estimate the aij based on the information oftheir indep endence/, whic ha l l o ws us to separate the t w oo riginal source signals s/1 /( t /) and s/2 /( t /) from theirmixtures x/1 /( t /) and x/2 /( t /) /. Fig/. /3 giv es the t w os ignals estimated b yt he ICA metho d/. As can b e seen/, theseare v ery close to the original source signals /(their signs are rev ersed/, but this has no signi/cance/./)Indep enden t comp onen ta nalysis w as originally dev elop ed to deal with problems that are closely relatedto the co c ktail/-part yp r o blem/. Since the recen ti ncrease of in terest in ICA/, it has b ecome clear that thisprinciple has a lot of other in teresting applications as w ell/./1

2211.06738.pdf

arXiv:2211.06738v1 [cs.AI] 12 Nov 2022Formalizing the presumption of independence Paul Christiano, Eric Neyman, Mark Xu Alignment Research Center Abstract Mathematical proof aims to deliver confident conclusions, but a ver y similar process of deduction can be used to make uncertain estimates that are open t o revision. A key ingredient in such reasoning is the use of a “default” estimate of E[XY] =E[X]E[Y] in the absence of any specific information about the correlation between XandY, which we call the presumption of independence . Reasoning based on this heuristic is commonplace, intuitively compellin g, and often quite successful—but completely informal. In this paper we introduce the concept of a heuristic estimator as a potential formalization of this type of defeasible reasoning. We introduce a set of intuitively de sirable coherence properties for heuristic estimators that are not satisfied by any existing cand idates. Then we present our main open problem: is there a heuristic estimator that formalizes intu itively valid applications of the presumption of independence without also accepting spuriou s arguments? Many formally-specified questions are very hard to settle wi th proofs. There are famous examples like the twin prime conjecture, but also countless more mund ane examples like how quickly the temperature of a simulated room would change if the window we re opened. Even when we cannot prove a theorem, we can often deductively arrive at a reasonable best guess about the truth of a claim or the behavior of a system. We can ma ke probabilistic arguments about the structure of the primes to estimate the density of twin pr imes, or about small molecules moving randomly in order to estimate the rate of heat transfer. This reasoning requires making best guesses about quantiti es that we can’t calculate exactly. We can often do this using the presumption of independence : when trying to estimate E[XY] without any knowledge about the relationship between XandY, we can use E[X]E[Y] as a default guess rather than remaining completely agnostic. For example, we can provisionally treat “ xis prime” and “x+2 is prime”as independent, or treat the velocities of differe nt air molecules as uncorrelated. This principle is sufficient to make plausible estimates abou t a very wide range of mathematical quantities. But it is not clear how to formalize this kind of d efeasible reasoning, nor is it clear how to generalize our default guess to the situation where we hav e arbitrary partial information about howXandYare related. Heuristic reasoning using the presumption of independence is distinct from running experiments or Monte Carlo simulations. We are not merely observing a lot of twin primes and inferring that there are probably infinitely many of them, or running simula tions of a room and observing how quickly the temperature changes—we have found a good reason that our answer should be right unless there is additional structure that we’ve overlooked which changes the answer. 1

1805.00899.pdf

AI safety via debate Geoffrey Irving∗Paul Christiano OpenAIDario Amodei Abstract To make AI systems broadly useful for challenging real-world tasks, we need them to learn complexhumangoalsandpreferences. Oneapproachtospecifyingcomplexgoalsaskshumansto judge during training which agent behaviors are safe and useful, but this approach can fail if the task is too complicated for a human to directly judge. To help address this concern, we propose training agents via self play on a zero sum debategame. Given a question or proposed action, two agents take turns making short statements up to a limit, then a human judges which of the agents gave the most true, useful information. In an analogy to complexity theory, debate with optimal play can answer any question in PSPACE given polynomial time judges (direct judging answers only NPquestions). In practice, whether debate works involves empirical questions about humans and the tasks we want AIs to perform, plus theoretical questions about the meaning of AI alignment. We report results on an initial MNIST experiment where agents compete to convince a sparse classifier, boosting the classifier’s accuracy from 59.4% to 88.9% given 6 pixels and from 48.2% to 85.2% given 4 pixels. Finally, we discuss theoretical and practical aspects of the debate model, focusing on potential weaknesses as the model scales up, and we propose future human and computer experiments to test these properties. 1 Introduction Learning to align an agent’s actions with the values and preferences of humans is a key challenge in ensuring that advanced AI systems remain safe [Russell et al., 2016]. Subtle problems in alignment can lead to unexpected and potentially unsafe behavior [Amodei et al., 2016], and we expect this problem to get worse as systems become more capable. Alignment is a training-time problem: it is difficult to retroactively fix the behavior and incentives of trained unaligned agents. Alignment likely requires interaction with humans during training, but care is required in choosing the precise form of the interaction as supervising the agent may itself be a challenging cognitive task. For some tasks it is harder to bring behavior in line with human goals than for others. In simple cases, humans can directly demonstrate the behavior—this is the case of supervised learning or imitation learning, for example classifying an image or using a robotic gripper to pick up a block. For these tasks alignment with human preferences can in principle be achieved by imitating the human, and is implicit in existing ML approaches (although issues of bias in the training data still arise, see e.g. Mitchell and Shadlen [2018]). Taking a step up in alignment difficulty, some tasks are too difficult for a human to perform, but a human can still judge the quality of behavior or answers once shown to them—for example a robot doing a backflip in an unnatural action space. This is the case of human preference-based reinforcement learning [Christiano et al., 2017]. We can make ∗Corresponding author: irving@openai.com 1arXiv:1805.00899v2 [stat.ML] 22 Oct 2018

2401.10020.pdf

Self-Rewarding Language Models Weizhe Yuan1,2Richard Yuanzhe Pang1,2Kyunghyun Cho2 Xian Li1Sainbayar Sukhbaatar1Jing Xu1Jason Weston1,2 1Meta2NYU Abstract We posit that to achieve superhuman agents, future models require super- human feedback in order to provide an adequate training signal. Current approaches commonly train reward models from human preferences, which may then be bottlenecked by human performance level, and secondly these separate frozen reward models cannot then learn to improve during LLM training. In this work, we study Self-Rewarding Language Models , where the language model itself is used via LLM-as-a-Judge prompting to provide its own rewards during training. We show that during Iterative DPO training that not only does instruction following ability improve, but also the ability to provide high-quality rewards to itself. Fine-tuning Llama 2 70B on three iterations of our approach yields a model that outperforms many existing systems on the AlpacaEval 2.0 leaderboard, including Claude 2, Gemini Pro, and GPT-4 0613. While there is much left still to explore, this work opens the door to the possibility of models that can continually improve in both axes. 1 Introduction Aligning Large Language Models (LLMs) using human preference data can vastly improve the instruction following performance of pretrained models [Ouyang et al., 2022, Bai et al., 2022a]. The standard approach of Reinforcement Learning from Human Feedback (RLHF) learns a reward model from these human preferences. The reward model is then frozen and used to train the LLM using RL, e.g., via PPO [Schulman et al., 2017]. A recent alternative is to avoid training the reward model at all, and directly use human preferences to train the LLM, as in Direct Preference Optimization [DPO; Rafailov et al., 2023]. In both cases, the approach is bottlenecked by the size and quality of the human preference data, and in the case of RLHF the quality of the frozen reward model trained from them as well. In this work, we instead propose to train a self-improving reward model that, rather than being frozen, is continually updating during LLM alignment, in order to avoid this bottleneck. The key to such an approach is to develop an agent that possesses all the abilities desired during training, rather than separating them out into distinct models such as a reward model and a language model. In the same way that pretraining and multitasking training of instruction following tasks allow task transfer by training on many tasks at once [Collobert and Weston, 2008, Radford et al., 2019, Ouyang et al., 2022], incorporating the reward model into that same system allows task transfer between the reward modeling task and the instruction following tasks. We thus introduce Self-Rewarding Language Models , that both (i) act as instruction following models generating responses for given prompts; and (ii) can generate and evaluate new instruction following examples to add to their own training set. We train these models using an Iterative DPO framework similar to that recently introduced in Xu et al. [2023].arXiv:2401.10020v2 [cs.CL] 8 Feb 2024

2401.12192.pdf

Text Embedding Inversion Attacks on Multilingual Language Models Yiyi Chen Heather Lent Johannes Bjerva Department of Computer Science, Aalborg University, Denmark {yiyic, hcle, jbjerva}@cs.aau.dk Abstract Representing textual information as real- numbered embeddings has become the norm in NLP. Moreover, with the rise of public interest in large language models (LLMs), Embeddings as a Service (EaaS) has rapidly gained traction as a business model. This is not without out- standing security risks, as previous research has demonstrated that sensitive data can be re- constructed from embeddings, even without knowledge of the underlying model that gen- erated them. However, such work is limited by its sole focus on English, leaving all other languages vulnerable to attacks by malicious actors. To this end, this work investigates LLM security from the perspective of multilingual embedding inversion. Concretely, we define the problem of black-box multilingual and cross- lingual inversion attacks, with special attention to a cross-domain scenario. Our findings re- veal that multilingual models are potentially more vulnerable to inversion attacks than their monolingual counterparts. This stems from the reduced data requirements for achieving comparable inversion performance in settings where the underlying language is not known a- priori. To our knowledge, this work is the first to delve into multilinguality within the context of inversion attacks, and our findings highlight the need for further investigation and enhanced defenses in the area of NLP Security. 1 Introduction Industrial applications of Natural Language Pro- cessing (NLP) typically utilize Large Language Models (LLMs) and frequently rely on vector databases via frameworks such as Embeddings as a Service (EaaS). In this context, rather than storing data as strings, high quality sentence embeddings are stored in a remote database instead. This allows end-users to efficiently search across these con- densed representations, which are seemingly im- pervious to privacy breaches. However, while such EaaS workflows have previously been assumed tobe secure, recent work has demonstrated that ac- cess to the embeddings is no more safe than raw text, as models can learn to decode these embed- dings (Song and Raghunathan, 2020; Morris et al., 2023; Zhou et al., 2023). As such, there is a sub- stantial threat to privacy if malicious actors are able to eavesdrop on communication channels between EaaS providors and customers, and access the em- beddings in the process. Decoding the content of these embeddings can be done via inversion attacks . After gaining access to embeddings and the black-box embedder via the EaaS API, the malicious actor can train an external model, which approximates the inversion function that reconstructs the text from the embeddings. Pre- vious work has proven has demonstrated that an exact match for data recreation can be obtained in specific settings, albeit with the limitation of assum- ing monolingual English models and embeddings (Morris et al., 2023). In a real-world scenario however, an eavesdrop- per may not necessarily know the language of the text encoded within the embedding. For instance, a Spanish EaaS provider might host its data in Ger- many, for a French-speaking company. Thus in this work we investigate three research questions: (i) To what extent are inversion attacks feasible in a multilingual setting?; (ii) Are attacks feasible and effective when the language is unknown a-priori?; (iii) Does cross-lingual transfer allow information to be leaked across the languages included in a multilingual model? Contributions In this work, we define the prob- lem of black-box multilingual and cross-lingual inversion attacks, with special attention to a cross- domain scenario. While previous research has suc- ceeded in reconstruction of tokens with bag-of- words approach (Song and Raghunathan, 2020) and sequences with informative words (Li et al., 2023), Morris et al. (2023) has proven the potentialarXiv:2401.12192v1 [cs.CL] 22 Jan 2024

2211.07793.pdf

EXTREME GENERATIVE IMAGE COMPRESSION BY LEARNING TEXT EMBEDDING FROM DIFFUSION MODELS A P REPRINT Zhihong Pan, Xin Zhou, Hao Tian Baidu Research (USA) ABSTRACT Transferring large amount of high resolution images over limited bandwidth is an important but very challenging task. Compressing images using extremely low bitrates (<0.1 bpp) has been studied but it often results in low quality images of heavy artifacts due to the strong constraint in the number of bits available for the compressed data. It is often said that a picture is worth a thousand words but on the other hand, language is very powerful in capturing the essence of an image using short descriptions. With the recent success of diffusion models for text-to-image generation, we propose a generative image compression method that demonstrates the potential of saving an image as a short text embedding which in turn can be used to generate high-fidelity images which is equivalent to the original one perceptually. For a given image, its corresponding text embedding is learned using the same optimization process as the text-to-image diffusion model itself, using a learnable text embedding as input after bypassing the original transformer. The optimization is applied together with a learning compression model to achieve extreme compression of low bitrates <0.1 bpp. Based on our experiments measured by a comprehensive set of image quality metrics, our method outperforms the other state-of-the-art deep learning methods in terms of both perceptual quality and diversity. 1 Introduction With the increasing amount of image streams available for broad range of applications, lossy image compression is a very useful technique for efficient image storage and transmission. Over the years, various engineered codes such as JPEG [ 30], JPEG2000 [ 52], and the more recent BPG[ 4] have been proposed to compress single images but their performance have saturated overall. More recently, deep learning based image compression methods have been studied [ 3,36,7]. These models are generally trained in an end-to-end fashion to minimize a rate-distortion object R+λD. HereRrepresents the entropy of latent representations which is estimated by an entropy model, Dis the difference between the original image and the compressed one, and λdetermines the desired trade-off between rate and distortion. When λis small, the optimization gives higher priority to compression rate so the resulted bitrate (evaluated as bits-per-pixel, bpp) is low. Consequently, the compressed image has lower quality due to higher Dloss term. With accuracy metrics like mean squared error (MSE) and multi-scale structural similarity (MS-SSIM) are often used for D, the low quality compressed images are usually blurry. For extremely low bitrates (<0.1 bpp), both engineered codecs and deep learning compression models are subject to very poor perceptual qualities. To tackle this problem, some recent methods [ 61,63,29,35] aim to restore less blurry image from highly compressed latent representations at the cost of accuracy. These model adopt generative adversarial networks (GAN) [ 19] to fully or partially replace the accuracy metrics in Dwith discrimination loss so they can generate sharp and realistic images even at very low bitrates. For the challenging task of extremely low bitrates, GAN is further exploited in more recent studies [ 2,11,25] to restore sharp images with minimized distortion and visual artifacts. However, they all inherit the drawback of unstable training from GAN, making it difficult to tune the training process for large datasets. In this paper, we propose the first generative image compression method with extremely low bitrates using denoising diffusion models. As it utilizes an existing text-to-image model which is already trained with a gigantic dataset, it is applicable to any type of image with no need of further tuning. Similar to GAN a few years back, denoising diffusion models [ 53,22,54] are gaining popularity increasingly for their advantages in generating images with high qualities in both fidelity and diversity without disadvantage of unstable training like GAN. In addition to unconditional image generation, diffusion models have also empowered the breakthrough developments in diffusion-based text-to-image generation models [ 47,38,43,49] which are able to createarXiv:2211.07793v1 [eess.IV] 14 Nov 2022

gu-dissertation-augmented.pdf

MODELING SEQUENCES WITH STRUCTURED STATE SPACES A DISSERTATION SUBMITTED TO THE DEPARTMENT OF DEPARTMENT OF COMPUTER SCIENCE AND THE COMMITTEE ON GRADUATE STUDIES OF STANFORD UNIVERSITY IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY Albert Gu June 2023

2108.05540.pdf

Unsupervised Corpus Aware Language Model Pre-training for Dense Passage Retrieval Luyu Gao and Jamie Callan Language Technologies Institute Carnegie Mellon University {luyug, callan}@cs.cmu.edu Abstract Recent research demonstrates the effective- ness of using fine-tuned language mod- els (LM) for dense retrieval. However, dense retrievers are hard to train, typically requiring heavily engineered fine-tuning pipelines to re- alize their full potential. In this paper, we iden- tify and address two underlying problems of dense retrievers: i) fragility to training data noise and ii) requiring large batches to robustly learn the embedding space. We use the re- cently proposed Condenser pre-training archi- tecture, which learns to condense information into the dense vector through LM pre-training. On top of it, we propose coCondenser, which adds an unsupervised corpus-level contrastive loss to warm up the passage embedding space. Retrieval experiments on MS-MARCO, Natu- ral Question, and Trivia QA datasets show that coCondenser removes the need for heavy data engineering such as augmentation, synthesis, or filtering, as well as the need for large batch training. It shows comparable performance to RocketQA, a state-of-the-art, heavily engi- neered system, using simple small batch fine- tuning.1 1 Introduction Building upon the advancements of pre-trained lan- guage models (LM; Devlin et al. (2019); Liu et al. (2019)), dense retrieval has become an effective paradigm for text retrieval (Lee et al., 2019; Chang et al., 2020; Karpukhin et al., 2020; Qu et al., 2021). Recent research has however found that fine-tuning dense retrievers to realize their capacity requires carefully designed fine-tuning techniques. Early works include iterative negative mining (Xiong et al., 2021) and multi-vector representations (Luan et al., 2020). The recent RocketQA system (Qu et al., 2021) significantly improves the performance 1Our code is available at https://github.com/ luyug/Condenserof a dense retriever by designing an optimized fine- tuning pipeline that includes i) denoising hard neg- atives, which corrects mislabeling, and ii) large batch training. While this is very effective, the en- tire pipeline is very heavy in computation and not feasible for people who do not have tremendous hardware resources, especially those in academia. In this paper, we ask, instead of directly using the pipeline, can we take the insights of RocketQA to perform language model pre-training such that the pre-trained model can be easily fine-tuned on any target query set. Concretely, we ask what the optimized training in RocketQA solves. We hypothesize that typi- cal LMs are sensitive to mislabeling, which can cause detrimental updates to the model weights. Denoising can effectively remove the bad samples and their updates. On the other hand, for most LMs, the CLS vectors are either trained with a simple task (Devlin et al., 2019) or not explicitly trained at all (Liu et al., 2019). These vectors are far from being able to form an embedding space of passages (Lee et al., 2019). The large training batches in RocketQA help the LM to stably learn to form the full embedding space. To this end, we want to pre-train an LM such that it is locally noise-resistant and has a well-structured global em- bedding space. For noise resistance, we borrow the Condenser pre-training architecture (Gao and Callan, 2021), which performs language model pre- training actively conditioned on the CLS vector. It produces an information-rich CLS representation that can robustly condense an input sequence. We then introduce a simple corpus level contrastive learning objective: given a target corpus of docu- ments to retrieve from, at each training step sample text span pairs from a batch of documents and train the model such that the CLS embeddings of two spans from the same document are close and spans from different documents are far apart. Combin- ing the two, we propose coCondenser pre-training,arXiv:2108.05540v1 [cs.IR] 12 Aug 2021

1501.05014.pdf

Experimental Simulation of Closed Timelike Curves Martin Ringbauer1,2∗, Matthew A. Broome1,2, Casey R. Myers1, Andrew G. White1,2and Timothy C. Ralph2 1Centre for Engineered Quantum Systems,2Centre for Quantum Computer and Communication Technology, School of Mathematics and Physics, University of Queensland, Brisbane, QLD 4072, Australia Closed timelike curves are among the most controversial features of modern physics. As legitimate solutions to Einstein’s field equations, they allow for time travel, which instinctively seems para- doxical. However, in the quantum regime these paradoxes can be resolved leaving closed timelike curves consistent with relativity. The study of these systems therefore provides valuable insight into non-linearities and the emergence of causal structures in quantum mechanics—essential for any for- mulation of a quantum theory of gravity. Here we experimentally simulate the non-linear behaviour of a qubit interacting unitarily with an older version of itself, addressing some of the fascinating effects that arise in systems traversing a closed timelike curve. These include perfect discrimination of non-orthogonal states and, most intriguingly, the ability to distinguish nominally equivalent ways of preparing pure quantum states. Finally, we examine the dependence of these effects on the initial qubit state, the form of the unitary interaction, and the influence of decoherence. INTRODUCTION One aspect of general relativity that has long intrigued physicists is the relative ease with which one can find so- lutions to Einstein’s field equations that contain closed timelike curves (CTCs)—causal loops in space-time that return to the same point in space and time [1–3]. Driven by apparent inconsistencies—like the grandfa- ther paradox—there have been numerous efforts, such as Novikov’s self-consistency principle [4] to reconcile them or Hawking’s chronology protection conjecture [5], to dis- prove the existence of CTCs. While none of these clas- sical hypotheses could be verified so far, the situation is particularly interesting in the quantum realm. In his seminal 1991 paper Deutsch showed for quantum sys- tems traversing CTCs there always exist unique solu- tions, which do not allow superluminal signalling [6, 7]. Quantum mechanics therefore allows for causality viola- tion without paradoxes whilst remaining consistent with relativity. Advances in the field of Deutsch CTCs have shown some very surprising and counter-intuitive results, such as the solution of NP-complete problems in polynomial time [8], unambiguous discrimination of any set of non- orthogonal states [9], perfect universal quantum state cloning [10, 11] and the violation of Heisenberg’s uncer- tainty principle [12]. The extraordinary claims of what one could achieve given access to a quantum system traversing a CTC have been disputed in the literature, with critics pointing out apparent inconsistencies in the theory such as the information paradox or the linearity trap [13, 14]. However, it has been shown that the theory can be formulated in such a way that these inconsisten- cies are resolved [7, 15]. ∗Electronic address: m.ringbauer@uq.edu.auModern experimental quantum simulation allows one to ask meaningful questions that provide insights into the behaviour of complex quantum systems. Initial results have been obtained in various areas of quantum mechan- ics [16–18] and in particular in the field of relativistic quantum information [19–23]. This recent experimental success, coupled with the growing interest for the study of non-linear extensions to quantum mechanics, motivates the question of whether the fundamentally non-linear dy- namics and the unique behaviour arising from CTCs can be simulated experimentally. In this article we use photonic systems to simulate the quantum evolution through a Deutsch CTC. We demon- strate how the CTC-traversing qubit adapts to changes in the input state |ψ⟩, and unitary interaction Uto en- sure physical consistency according to Deutsch’s consis- tency relation [6]. We observe non-linear evolution in the circuit suggested by Bacon [8] and enhanced distin- guishability of two non-orthogonal states after the action of an optimised version of a circuit proposed by Brun et al. [9]. Using the self-consistent formulation of Ref. [7] we then move beyond the simplest implementations and find a striking difference in the behaviour of the system for direct as opposed to entanglement-assisted state prepa- ration. Finally, we explore the system’s sensitivity to decoherence. U U FIG. 1: Model of a quantum state |ψ⟩interacting with an older version of itself. This situation can equivalently be interpreted as a chronology-respecting qubit interacting with a qubit trapped in a CTC. The CTC in general consists of a causal worldline with its past and future ends connected via a wormhole (indicated by black triangles).arXiv:1501.05014v1 [quant-ph] 20 Jan 2015

2310.18168.pdf

PERSONAS AS A WAY TO MODEL TRUTHFULNESS IN LANGUAGE MODELS Nitish Joshi1∗Javier Rando2∗Abulhair Saparov1Najoung Kim3He He1 1New York University2ETH Zurich3Boston University {nitish}@nyu.edu {jrando}@ethz.ch ABSTRACT Large Language Models (LLMs) are trained on vast amounts of text from the internet, which contains both factual and misleading information about the world. Can language models discern truth from falsehood in this contradicting data? Expanding on the view that LLMs can model different agents producing the corpora, we hypothesize that they can cluster truthful text by modeling a truthful persona : a group of agents that are likely to produce truthful text and share similar features. For example, trustworthy sources like Wikipedia and Science usually use formal writing styles and make consistent claims. By modeling this persona, LLMs can generalize truthfulness beyond the specific contexts in which each agent generated the training text. For example, the model can infer that the agent “Wikipedia” will behave truthfully on topics that were only generated by “Science” because they share a persona. We first show evidence for the persona hypothesis via two observations: (1) we can probe whether a model’s answer will be truthful before it is generated; (2) finetuning a model on a set of facts improves its truthfulness on unseen topics. Next, using arithmetics as a synthetic environment, we show that language models can separate true and false statements, and generalize truthfulness across agents; but only if agents in the training data share a truthful generative process that enables the creation of a truthful persona. Overall, our findings suggest that models can exploit hierarchical structures in the data to learn abstract concepts like truthfulness. 1 I NTRODUCTION Large Language Models (LLMs) are pretrained on increasing amounts of data from the internet (Brown et al., 2020; Chowdhery et al., 2022)—a noisy, and mostly uncurated corpus—which contains both truthful statements about the world and untruthful statements such as misconceptions and conspiracy theories. The false claims in the data pose a risk of misinformation as they can be propogated by the model (Lin et al., 2021). Intriguingly, recent work shows that the truth value of a statement can be elicited from its embeddings (Burns et al., 2022; Li et al., 2023). This motivates the main question of this work: what mechanism do LLMs use to distinguish truth from falsehood despite noise in the data? Consider two contradicting statements: "people with type A blood are ambitious" (false) and "blood type does not imply any personality traits" (true). When asked about the relation between blood type and personality, the classic view of language models suggests that it will generate the most frequent statement, regardless of whether it is true. However, we observe that slight changes in the question can steer the model to produce any of the two (Figure 1). This suggests that frequency alone is not sufficient to explain model behavior. Andreas (2022) hypothesizes that LLMs can infer the agent who produced the context and generate continuations according to the agent’s goals and beliefs. In this example, given the question "What personality does someone with type A blood have?" with a false presupposition (Kim et al., 2022), the model may infer that the agent who asks the question believes that blood type influences personality, and thus generate an answer following this (false) belief. If the ∗equal contribution 1arXiv:2310.18168v2 [cs.CL] 30 Oct 2023

1712.03346.pdf

Variational auto-encoding of protein sequences Sam Sinai∗ Harvard University samsinai@g.harvard.eduEric Kelsic†‡ Harvard Medical School eric kelsic@hms.harvard.edu George M. Church§†‡ Harvard Medical School church labadmin@hms.harvard.eduMartin A. Nowak∗‡¶ Harvard University martin nowak@harvard.edu Abstract Proteins are responsible for the most diverse set of functions in biology. The abil- ity to extract information from protein sequences and to predict the effects of mu- tations is extremely valuable in many domains of biology and medicine. However the mapping between protein sequence and function is complex and poorly under- stood. Here we present an embedding of natural protein sequences using a Vari- ational Auto-Encoder and use it to predict how mutations affect protein function. We use this unsupervised approach to cluster natural variants and learn interac- tions between sets of positions within a protein. This approach generally performs better than baseline methods that consider no interactions within sequences, and in some cases better than the state-of-the-art approaches that use the inverse-Potts model. This generative model can be used to computationally guide exploration of protein sequence space and to better inform rational and automatic protein design. 1 Introduction Protein engineering is of increasing importance in modern therapeutics. Designing novel proteins that perform a particular function is challenging as the number of functional proteins compared to all possible protein sequences is miniscule. This renders naive experimental search for desirable variants intractable. Hence, a computational heuristic that can narrow the experimental search space (virtual screening) is extremely valuable. While a variety of energy-based models for protein folding have been used in the past decades, re- cent advances in machine learning, particularly in the domain of generative models, have opened up new avenues for computational protein design. Rich databases of protein sequences that document functional proteins found in living organisms provide us with ample training data. The majority of these datasets lack labels (indicators of their performance) however, which prompts for an un- supervised learning approach. As these sequences arise from closely related living organisms, it is reasonable to assume that they are functional (and also similar in their functionality). Given the sparse, unstructured, and discrete space that protein sequences exist in, it is prudent to anchor the search for functional sequences on a known protein with the desired functionality. Start- ing from that sequence of interest, we can search public databases of sequence variants from related ∗Program for Evolutionary Dynamics, Department of Organismic and Evolutionary Biology †Wyss Institute ‡To whom correspondence should be directed §Department of Genetics ¶Department of Mathematics 1arXiv:1712.03346v3 [q-bio.QM] 3 Jan 2018

2309.16797.pdf

PROMPTBREEDER : SELF-REFERENTIAL SELF-IMPROVEMENT VIAPROMPT EVOLUTION Chrisantha Fernando, Dylan Banarse, Henryk Michalewski, Simon Osindero, Tim Rockt ¨aschel Google DeepMind {chrisantha,dylski,henrykm,osindero,rocktaschel }@google.com ABSTRACT Popular prompt strategies like Chain-of-Thought Prompting can dramatically im- prove the reasoning abilities of Large Language Models (LLMs) in various do- mains. However, such hand-crafted prompt-strategies are often sub-optimal. In this paper, we present P ROMPTBREEDER , a general-purpose self-referential self- improvement mechanism that evolves and adapts prompts for a given domain. Driven by an LLM, Promptbreeder mutates a population of task-prompts, evalu- ates them for fitness on a training set, and repeats this process over multiple gen- erations to evolve task-prompts. Crucially, the mutation of these task-prompts is governed by mutation-prompts that the LLM generates and improves throughout evolution in a self-referential way. That is, Promptbreeder is not just improving task-prompts, but it is also improving the mutation-prompts that improve these task-prompts. Promptbreeder outperforms state-of-the-art prompt strategies such as Chain-of-Thought and Plan-and-Solve Prompting on commonly used arith- metic and commonsense reasoning benchmarks. Furthermore, Promptbreeder is able to evolve intricate task-prompts for the challenging problem of hate speech classification. 1 I NTRODUCTION Prompting is central to the downstream performance of foundation models. For example, different prompt strategies1can have a significant impact on a model’s reasoning abilities (Wei et al., 2022; Nye et al., 2021; Zhou et al., 2022; Wang et al., 2022; Zhou et al., 2023; Wang et al., 2023b), multi- modal processing abilities (Yang et al., 2023b; Wang et al., 2023d), or tool use abilities (Yao et al., 2022; Schick et al., 2023). Furthermore, prompting can improve model distillation (Wang et al., 2023c; Hsieh et al., 2023) and it can be used to simulate agentic behavior (Wang et al., 2023a; Park et al., 2023; Wu et al., 2023). However, these prompt strategies are manually engineered. Since the specific way a prompt is phrased can have a dramatic effect on its utility (Madaan & Yazdanbakhsh, 2022), it raises the question of whether prompt engineering can be automated. Automatic Prompt Engineer (APE, Zhou et al., 2023) attempts to address this by generating an initial distribution of prompts using another prompt that infers the problem from a number of input-output examples from the dataset. However, Zhou et al. found “diminishing returns to further selection rounds as the qual- ity seems to stabilize after three rounds”, and consequently abandoned the use of an iterative APE. We propose a solution to the problem of diminishing returns via a diversity maintaining evolutionary algorithm for self-referential self-improvement of prompts for LLMs. Schmidhuber (1990) notes that the “program of a neural network is its weight matrix”. Con- sequently, this “program” can be changed in a self-referential way by the neural network it- self (Schmidhuber, 1993; Irie et al., 2022). Such a neural network that improves itself, as well as improving the way it improves itself, might be an important stepping stone towards open-ended self-referential self-improvement of AIs (Schmidhuber, 2003). However, self-improvement via self- referential weight matrices is costly as it requires additional parameters that modify all of the model’s 1See Appendix A for definitions of terminology. 1arXiv:2309.16797v1 [cs.CL] 28 Sep 2023

2404.12253v1.pdf

Toward Self-Improvement of LLMs via Imagination, Searching, and Criticizing Ye Tian∗, Baolin Peng∗, Linfeng Song∗, Lifeng Jin, Dian Yu, Haitao Mi†, Dong Yu Tencent AI Lab, Bellevue, WA {yaptian,baolinpeng,lfsong,lifengjin,yudian,haitaomi}@global.tencent.com Abstract Despite the impressive capabilities of Large Language Models (LLMs) on various tasks, they still struggle with scenarios that involves complex reasoning and plan- ning. Recent work proposed advanced prompting techniques and the necessity of fine-tuning with high-quality data to augment LLMs’ reasoning abilities. However, these approaches are inherently constrained by data availability and quality. In light of this, self-correction and self-learning emerge as viable solutions, employing strategies that allow LLMs to refine their outputs and learn from self-assessed rewards. Yet, the efficacy of LLMs in self-refining its response, particularly in complex reasoning and planning task, remains dubious. In this paper, we introduce ALPHA LLM for the self-improvements of LLMs, which integrates Monte Carlo Tree Search (MCTS) with LLMs to establish a self-improving loop, thereby enhanc- ing the capabilities of LLMs without additional annotations. Drawing inspiration from the success of AlphaGo, ALPHA LLM addresses the unique challenges of combining MCTS with LLM for self-improvement, including data scarcity, the vastness search spaces of language tasks, and the subjective nature of feedback in language tasks. ALPHA LLM is comprised of prompt synthesis component, an efficient MCTS approach tailored for language tasks, and a trio of critic models for precise feedback. Our experimental results in mathematical reasoning tasks demon- strate that ALPHA LLM significantly enhances the performance of LLMs without additional annotations, showing the potential for self-improvement in LLMs. 1 Introduction LLMs, trained on trillions of tokens with billions of parameters have shown unparalleled capabilities in a wide range of natural language processing tasks (Touvron et al., 2023b; Team et al., 2023; OpenAI, 2023). Nevertheless, they continue to face challenges in scenarios requiring complex reasoning and strategic planning (Valmeekam et al., 2022; Stechly et al., 2024). While advanced prompting approaches such as Chain, Tree, Graph-of-Thought (Wei et al., 2022; Yao et al., 2024; Besta et al., 2024; Ding et al., 2023), which generate intermediate steps in the reasoning process demonstrate large improvements on reasoning capability of LLMs, it remains essential to fine-tune LLMs using a substantial volume of high-quality, supervised data to fundamentally improve the model performance (Nye et al., 2021; Lewkowycz et al., 2022; Chung et al., 2022). This methodology is inherently limited by the scope and quality of data that humans can provide. Considering existing challenges, the concept of self-correction and self-learning have been proposed as promising solutions (Madaan et al., 2024; Saunders et al., 2022; Chen et al., 2024). Within these framework, LLMs typically operate by employing two main strategies: 1) they continuously refine their responses based on the feedback of their past responses, and 2) they extensively sample ∗Equal Contribution; †Corresponding Author Work in progress.arXiv:2404.12253v1 [cs.CL] 18 Apr 2024

2005.10242.pdf

Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere Tongzhou Wang1Phillip Isola1 Abstract Contrastive representation learning has been out- standingly successful in practice. In this work, we identify two key properties related to the con- trastive loss: (1) alignment (closeness) of features from positive pairs, and (2) uniformity of the in- duced distribution of the (normalized) features on the hypersphere. We prove that, asymptotically, the contrastive loss optimizes these properties, and analyze their positive effects on downstream tasks. Empirically, we introduce an optimizable metric to quantify each property. Extensive exper- iments on standard vision and language datasets confirm the strong agreement between both met- rics and downstream task performance. Directly optimizing for these two metrics leads to repre- sentations with comparable or better performance at downstream tasks than contrastive learning. Project Page: ssnl.github.io/hypersphere . Code: github.com/SsnL/align uniform . github.com/SsnL/moco align uniform . 1. Introduction A vast number of recent empirical works learn representa- tions with a unit ℓ2norm constraint, effectively restricting the output space to the unit hypersphere (Parkhi et al., 2015; Schroff et al., 2015; Liu et al., 2017; Hasnat et al., 2017; Wang et al., 2017; Bojanowski & Joulin, 2017; Mettes et al., 2019; Hou et al., 2019; Davidson et al., 2018; Xu & Durrett, 2018), including many unsupervised contrastive represen- tation learning methods (Wu et al., 2018; Bachman et al., 2019; Tian et al., 2019; He et al., 2019; Chen et al., 2020a). Intuitively, having the features live on the unit hypersphere leads to several desirable traits. Fixed-norm vectors are known to improve training stability in modern machine learning where dot products are ubiquitous (Xu & Durrett, 1MIT Computer Science & Artificial Intelligence Lab (CSAIL). Correspondence to: Tongzhou Wang <tongzhou@mit.edu >. Proceedings of the 37thInternational Conference on Machine Learning , Online, PMLR 119, 2020. Copyright 2020 by the au- thor(s). Alignment:Similar samples have similar featuresAlignment: Similar samples have similar features. (Figure inspired by Tian et al. (2019).) Feature Density Uniformity: Preserve maximal information Uniformity: Preserve maximal information. Figure 1: Illustration of alignment and uniformity of fea- ture distributions on the output unit hypersphere. STL-10 (Coates et al., 2011) images are used for demonstration. 2018; Wang et al., 2017). Moreover, if features of a class are sufficiently well clustered, they are linearly separable with the rest of feature space (see Figure 2), a common criterion used to evaluate representation quality. While the unit hypersphere is a popular choice of feature space, not all encoders that map onto it are created equal. Recent works argue that representations should addition- ally be invariant to unnecessary details, and preserve as much information as possible (Oord et al., 2018; Tian et al., 2019; Hjelm et al., 2018; Bachman et al., 2019). Let us call these two properties alignment anduniformity (see Figure 1). Alignment favors encoders that assign similararXiv:2005.10242v10 [cs.LG] 15 Aug 2022

Improving-Memory-Search-through-Model-Based-Cue-Selection.pdf

IMPROVING MEMORY SEARCH 1 . Improving Memory Search through Model-Based Cue Selection Charlotte A. Cornell1, Kenneth A. Norman2, Thomas L. Griffiths2,3, and Qiong Zhang1,4 1Psychology Department, Rutgers University–New Brunswick 2Psychology Department, Princeton University 3Computer Science Department, Princeton University 4Computer Science Department, Rutgers University–New Brunswick Author Note This work was supported by a start-up fund awarded to Q.Z. by Rutgers University–New Brunsiwck and the National Science Foundation (BCS-2316716) awarded to Q.Z.. Correspondence concerning this article should be addressed to Qiong Zhang <qiong.z@rutgers.edu>

tr00-004.pdf

/CC /D6/CP/CX/D2/CX/D2/CV /C8/D6/D3 /CS/D9/CR/D8/D7 /D3/CU /BX/DC/D4 /CT/D6/D8/D7 /CQ /DD /C5/CX/D2/CX/D1/CX/DE/CX/D2/CV /BV/D3/D2 /D8/D6/CP/D7/D8/CX/DA /CT/BW/CX/DA /CT/D6/CV/CT/D2/CR/CT/BZ/BV/C6/CD /CC/CA /BE/BC/BC/BC/B9/BC/BC/BG/BZ/CT/D3/AB/D6/CT/DD /BX/BA /C0/CX/D2 /D8/D3/D2/BZ/CP/D8/D7/CQ /DD /BV/D3/D1/D4/D9/D8/CP/D8/CX/D3/D2/CP/D0 /C6/CT/D9/D6/D3/D7/CR/CX/CT/D2/CR/CT /CD/D2/CX/D8/CD/D2/CX/DA /CT/D6/D7/CX/D8 /DD /BV/D3/D0/D0/CT/CV/CT /C4/D3/D2/CS/D3/D2/BD/BJ /C9/D9/CT/CT/D2 /CB/D5/D9/CP/D6/CT/B8 /C4/D3/D2/CS/D3/D2 /CF /BV/BD/C6 /BF/BT/CA/B8 /CD/BA/C3/BA/CW/D8/D8/D4/BM/BB/BB/DB/DB/DB/BA/CV/CP/D8/D7/CQ/DD /BA/D9 /CR/D0/BA /CP/CR /BA/D9/CZ /BB/BT/CQ/D7/D8/D6/CP/CR/D8/C1/D8 /CX/D7 /D4 /D3/D7/D7/CX/CQ/D0/CT /D8/D3 /CR/D3/D1 /CQ/CX/D2/CT /D1 /D9/D0/D8/CX/D4/D0/CT /D4/D6/D3/CQ/CP/CQ/CX/D0/CX/D7/D8/CX/CR /D1/D3 /CS/CT/D0/D7 /D3/CU /D8/CW/CT /D7/CP/D1/CT /CS/CP/D8/CP /CQ /DD/D1 /D9/D0/D8/CX/D4/D0/DD/CX/D2/CV/D8/CW/CT/CX/D6 /D4/D6/D3/CQ/CP/CQ/CX/D0/CX/D8 /DD /CS/CX/D7/D8/D6/CX/CQ/D9/D8/CX/D3/D2/D7 /D8/D3/CV/CT/D8/CW/CT/D6 /CP/D2/CS /D8/CW/CT/D2 /D6/CT/D2/D3/D6/D1/CP/D0/CX/DE/CX/D2/CV/BA /CC/CW/CX/D7 /CX/D7 /CP /DA /CT/D6/DD /CTÆ/CR/CX/CT/D2 /D8/DB /CP /DD /D8/D3 /D1/D3 /CS/CT/D0 /CW/CX/CV/CW/B9/CS/CX/D1/CT/D2/D7/CX/D3/D2/CP/D0 /CS/CP/D8/CP /DB/CW/CX/CR /CW /D7/CX/D1 /D9/D0/D8/CP/D2/CT/D3/D9/D7/D0/DD /D7/CP/D8/CX/D7/AC/CT/D7 /D1/CP/D2 /DD /CS/CX/AB/CT/D6/CT/D2 /D8 /D0/D3 /DB/B9/CS/CX/D1/CT/D2/D7/CX/D3/D2/CP/D0 /CR/D3/D2/D7/D8/D6/CP/CX/D2 /D8/D7 /CQ /CT/CR/CP/D9/D7/CT /CT/CP/CR /CW /CX/D2/CS/CX/DA/CX/CS/D9/CP/D0 /CT/DC/D4 /CT/D6/D8 /D1/D3 /CS/CT/D0 /CR/CP/D2 /CU/D3 /CR/D9/D7 /D3/D2 /CV/CX/DA/CX/D2/CV /CW/CX/CV/CW/D4/D6/D3/CQ/CP/CQ/CX/D0/CX/D8 /DD/D8 /D3/CS /CP /D8 /CP/DA /CT/CR/D8/D3/D6/D7 /D8/CW/CP/D8 /D7/CP/D8/CX/D7/CU/DD /CY/D9/D7/D8 /D3/D2/CT /D3/CU /D8/CW/CT /CR/D3/D2/D7/D8/D6/CP/CX/D2 /D8/D7/BA /BW/CP/D8/CP /DA /CT/CR/D8/D3/D6/D7 /D8/CW/CP/D8 /D7/CP/D8/CX/D7/CU/DD/D8/CW/CX/D7 /D3/D2/CT /CR/D3/D2/D7/D8/D6/CP/CX/D2 /D8 /CQ/D9/D8 /DA/CX/D3/D0/CP/D8/CT /D3/D8/CW/CT/D6 /CR/D3/D2/D7/D8/D6/CP/CX/D2 /D8/D7 /DB/CX/D0/D0 /CQ /CT /D6/D9/D0/CT/CS /D3/D9/D8 /CQ /DD /D8/CW/CT/CX/D6 /D0/D3 /DB /D4/D6/D3/CQ/CP/CQ/CX/D0/CX/D8 /DD/D9/D2/CS/CT/D6 /D8/CW/CT /D3/D8/CW/CT/D6 /CT/DC/D4 /CT/D6/D8/D7/BA /CC /D6/CP/CX/D2/CX/D2/CV /CP /D4/D6/D3 /CS/D9/CR/D8 /D3/CU /CT/DC/D4 /CT/D6/D8/D7 /CP/D4/D4 /CT/CP/D6/D7 /CS/CXÆ/CR/D9/D0/D8 /CQ /CT/CR/CP/D9/D7/CT/B8 /CX/D2 /CP/CS/CS/CX/D8/CX/D3/D2/D8/D3 /D1/CP/DC/CX/D1/CX/DE/CX/D2/CV /D8/CW/CT /D4/D6/D3/CQ/CP/CQ/CX/D0/CX/D8 /DD /D8/CW/CP/D8 /CT/CP/CR /CW /CX/D2/CS/CX/DA/CX/CS/D9/CP/D0 /CT/DC/D4 /CT/D6/D8 /CP/D7/D7/CX/CV/D2/D7 /D8/D3 /D8/CW/CT /D3/CQ/D7/CT/D6/DA /CT/CS /CS/CP/D8/CP/B8 /CX/D8/CX/D7 /D2/CT/CR/CT/D7/D7/CP/D6/DD /D8/D3 /D1/CP/CZ /CT /D8/CW/CT /CT/DC/D4 /CT/D6/D8/D7 /CQ /CT /CP/D7 /CS/CX/AB/CT/D6/CT/D2 /D8 /CP/D7 /D4 /D3/D7/D7/CX/CQ/D0/CT/BA /CC/CW/CX/D7 /CT/D2/D7/D9/D6/CT/D7 /D8/CW/CP/D8 /D8/CW/CT /D4/D6/D3 /CS/D9/CR/D8/D3/CU /D8/CW/CT/CX/D6 /CS/CX/D7/D8/D6/CX/CQ/D9/D8/CX/D3/D2/D7 /CX/D7 /D7/D1/CP/D0/D0 /DB/CW/CX/CR /CW/CP /D0 /D0 /D3 /DB/D7 /D8/CW/CT /D6/CT/D2/D3/D6/D1/CP/D0/CX/DE/CP/D8/CX/D3/D2 /D8/D3 /D1/CP/CV/D2/CX/CU/DD /D8/CW/CT /D4/D6/D3/CQ/CP/CQ/CX/D0/CX/D8 /DD/D3/CU /D8/CW/CT /CS/CP/D8/CP /D9/D2/CS/CT/D6 /D8/CW/CT /D4/D6/D3 /CS/D9/CR/D8 /D3/CU /CT/DC/D4 /CT/D6/D8/D7 /D1/D3 /CS/CT/D0/BA /BY /D3/D6/D8/D9/D2/CP/D8/CT/D0/DD /B8 /CX/CU /D8/CW/CT /CX/D2/CS/CX/DA/CX/CS/D9/CP/D0 /CT/DC/D4 /CT/D6/D8/D7 /CP/D6/CT/D8/D6/CP/CR/D8/CP/CQ/D0/CT /D8/CW/CT/D6/CT /CX/D7 /CP/D2 /CTÆ/CR/CX/CT/D2 /D8/DB /CP /DD /D8/D3 /D8/D6/CP/CX/D2 /CP /D4/D6/D3 /CS/D9/CR/D8 /D3/CU /CT/DC/D4 /CT/D6/D8/D7/BA/BD /C1/D2 /D8/D6/D3 /CS/D9/CR/D8/CX/D3/D2/C7/D2/CT /DB /CP /DD /D3/CU /D1/D3 /CS/CT/D0/CX/D2/CV /CP /CR/D3/D1/D4/D0/CX/CR/CP/D8/CT/CS/B8 /CW/CX/CV/CW/B9/CS/CX/D1/CT/D2/D7/CX/D3/D2/CP/D0 /CS/CP/D8/CP /CS/CX/D7/D8/D6/CX/CQ/D9/D8/CX/D3/D2 /CX/D7 /D8/D3 /D9/D7/CT /CP /D0/CP/D6/CV/CT /D2 /D9/D1 /CQ/CT /D6/D3/CU /D6/CT/D0/CP/D8/CX/DA /CT/D0/DD /D7/CX/D1/D4/D0/CT /D4/D6/D3/CQ/CP/CQ/CX/D0/CX/D7/D8/CX/CR /D1/D3 /CS/CT/D0/D7 /CP/D2/CS /D8/D3 /D7/D3/D1/CT/CW/D3 /DB/CR /D3 /D1 /CQ/CX/D2/CT /D8/CW/CT /CS/CX/D7/D8/D6/CX/CQ/D9/D8/CX/D3/D2/D7 /D7/D4 /CT/CR/CX/AC/CT/CS /CQ /DD/CT/CP/CR /CW/D1 /D3 /CS /CT /D0 /BA /BT /DB /CT/D0/D0/B9/CZ/D2/D3 /DB/D2 /CT/DC/CP/D1/D4/D0/CT /D3/CU /D8/CW/CX/D7 /CP/D4/D4/D6/D3/CP/CR /CW/CX /D7 /CP /D1/CX/DC/D8/D9/D6/CT /D3/CU /BZ/CP/D9/D7/D7/CX/CP/D2/D7 /CX/D2 /DB/CW/CX/CR /CW /CT/CP/CR /CW/D7/CX/D1/D4/D0/CT /D1/D3 /CS/CT/D0 /CX/D7 /CP /BZ/CP/D9/D7/D7/CX/CP/D2 /CP/D2/CS /D8/CW/CT /CR/D3/D1 /CQ/CX/D2/CP/D8/CX/D3/D2 /D6/D9/D0/CT /CR/D3/D2/D7/CX/D7/D8/D7 /D3/CU /D8/CP/CZ/CX/D2/CV /CP /DB /CT/CX/CV/CW /D8/CT/CS /CP/D6/CX/D8/CW/D1/CT/D8/CX/CR/D1/CT/CP/D2 /D3/CU /D8/CW/CT /CX/D2/CS/CX/DA/CX/CS/D9/CP/D0 /CS/CX/D7/D8/D6/CX/CQ/D9/D8/CX/D3/D2/D7/BA /CC/CW/CX/D7 /CX/D7 /CT/D5/D9/CX/DA /CP/D0/CT/D2 /D8/D8 /D3/CP /D7 /D7 /D9 /D1 /CX /D2 /CV /CP /D2/D3 /DA /CT/D6/CP/D0/D0 /CV/CT/D2/CT/D6/CP/D8/CX/DA /CT/D1 /D3 /CS /CT /D0/CX/D2 /DB/CW/CX/CR /CW/CT /CP /CR /CW/CS /CP /D8 /CP/DA /CT/CR/D8/D3/D6 /CX/D7 /CV/CT/D2/CT/D6/CP/D8/CT/CS /CQ /DD/AC /D6 /D7 /D8 /CR /CW/D3 /D3/D7/CX/D2/CV /D3/D2/CT /D3/CU /D8/CW/CT /CX/D2/CS/CX/DA/CX/CS/D9/CP/D0 /CV/CT/D2/CT/D6/CP/D8/CX/DA /CT /D1/D3 /CS/CT/D0/D7/CP/D2/CS /D8/CW/CT/D2 /CP/D0/D0/D3 /DB/CX/D2/CV /D8/CW/CP/D8 /CX/D2/CS/CX/DA/CX/CS/D9/CP/D0 /D1/D3 /CS/CT/D0 /D8/D3 /CV/CT/D2/CT/D6/CP/D8/CT /D8/CW/CT /CS/CP/D8/CP /DA /CT/CR/D8/D3/D6/BA /BV/D3/D1 /CQ/CX/D2/CX/D2/CV /D1/D3 /CS/CT/D0/D7 /CQ /DD/CU/D3/D6/D1/CX/D2/CV /CP /D1/CX/DC/D8/D9/D6/CT /CX/D7 /CP/D8/D8/D6/CP/CR/D8/CX/DA /CT /CU/D3/D6 /D7/CT/DA /CT/D6/CP/D0 /D6/CT/CP/D7/D3/D2/D7/BA /C1/D8 /CX/D7 /CT/CP/D7/DD /D8/D3 /AC/D8 /D1/CX/DC/D8/D9/D6/CT/D7 /D3/CU /D8/D6/CP/CR/D8/CP/CQ/D0/CT /D1/D3 /CS/CT/D0/D7/D8/D3 /CS/CP/D8/CP /D9/D7/CX/D2/CV /BX/C5 /D3/D6 /CV/D6/CP/CS/CX/CT/D2 /D8 /CP/D7/CR/CT/D2 /D8 /CP/D2/CS/B8 /CX/CU /D8/CW/CT /CX/D2/CS/CX/DA/CX/CS/D9/CP/D0 /D1/D3 /CS/CT/D0/D7 /CS/CX/AB/CT/D6 /CP/D0 /D3 /D8 /B8 /D8/CW/CT /D1/CX/DC/D8/D9/D6/CT /CX/D7/D0/CX/CZ /CT/D0/DD /D8/D3 /CQ/CT /CP /CQ /CT/D8/D8/CT/D6 /AC/D8 /D8/D3 /D8/CW/CT /D8/D6/D9/CT /CS/CX/D7/D8/D6/CX/CQ/D9/D8/CX/D3/D2 /D3/CU /D8/CW/CT /CS/CP/D8/CP /D8/CW/CP/D2 /CP /D6/CP/D2/CS/D3/D1 /CR /CW/D3/CX/CR/CT /CP/D1/D3/D2/CV /D8/CW/CT/CX/D2/CS/CX/DA/CX/CS/D9/CP/D0 /D1/D3 /CS/CT/D0/D7/BA /C1/D2/CS/CT/CT/CS/B8 /CX/CU /D7/D9Æ/CR/CX/CT/D2 /D8/D0/DD /D1/CP/D2 /DD /D1/D3 /CS/CT/D0/D7 /CP/D6/CT /CX/D2/CR/D0/D9/CS/CT/CS /CX/D2 /D8/CW/CT /D1/CX/DC/D8/D9/D6/CT/B8 /CX/D8 /CX/D7 /D4 /D3/D7/D7/CX/CQ/D0/CT/D8/D3 /CP/D4/D4/D6/D3 /DC/CX/D1/CP/D8/CT /CR/D3/D1/D4/D0/CX/CR/CP/D8/CT/CS /D7/D1/D3 /D3/D8/CW /CS/CX/D7/D8/D6/CX/CQ/D9/D8/CX/D3/D2/D7 /CP/D6/CQ/CX/D8/D6/CP/D6/CX/D0/DD /CP/CR/CR/D9/D6/CP/D8/CT/D0/DD /BA/CD/D2/CU/D3/D6/D8/D9/D2/CP/D8/CT/D0/DD /B8 /D1/CX/DC/D8/D9/D6/CT /D1/D3 /CS/CT/D0/D7 /CP/D6/CT /DA /CT/D6/DD /CX/D2/CTÆ/CR/CX/CT/D2 /D8 /CX/D2 /CW/CX/CV/CW/B9/CS/CX/D1/CT/D2/D7/CX/D3/D2/CP/D0 /D7/D4/CP/CR/CT/D7/BA /BV/D3/D2/D7/CX/CS/CT/D6/B8 /CU/D3/D6/CT/DC/CP/D1/D4/D0/CT/B8 /D8/CW/CT /D1/CP/D2/CX/CU/D3/D0/CS /D3/CU /CU/CP/CR/CT /CX/D1/CP/CV/CT/D7/BA /C1/D8 /D8/CP/CZ /CT/D7 /CP/CQ /D3/D9/D8 /BF/BH /D6/CT/CP/D0 /D2 /D9/D1 /CQ /CT/D6/D7 /D8/D3 /D7/D4 /CT/CR/CX/CU/DD /D8/CW/CT /D7/CW/CP/D4 /CT/B8/D4 /D3/D7/CT/B8 /CT/DC/D4/D6/CT/D7/D7/CX/D3/D2 /CP/D2/CS /CX/D0/D0/D9/D1/CX/D2/CP/D8/CX/D3/D2 /D3/CU /CP /CU/CP/CR/CT /CP/D2/CS/B8 /D9/D2/CS/CT/D6 /CV/D3 /D3 /CS /DA/CX/CT/DB/CX/D2/CV /CR/D3/D2/CS/CX/D8/CX/D3/D2/D7/B8 /D3/D9/D6 /D4 /CT/D6/CR/CT/D4/D8/D9/CP/D0/D7/DD/D7/D8/CT/D1/D7 /D4/D6/D3 /CS/D9/CR/CT /CP /D7/CW/CP/D6/D4 /D4 /D3/D7/D8/CT/D6/CX/D3/D6 /CS/CX/D7/D8/D6/CX/CQ/D9/D8/CX/D3/D2 /D3/D2 /D8/CW/CX/D7 /BF/BH/B9/CS/CX/D1/CT/D2/D7/CX/D3/D2/CP/D0 /D1/CP/D2/CX/CU/D3/D0/CS/BA /CC/CW/CX/D7 /CR/CP/D2/D2/D3/D8/CQ /CT /CS/D3/D2/CT /D9/D7/CX/D2/CV /CP /D1/CX/DC/D8/D9/D6/CT /D3/CU /D1/D3 /CS/CT/D0/D7 /CT/CP/CR /CW/D3 /CU/DB /CW /CX /CR /CW /CX/D7 /D8/D9/D2/CT/CS /CX/D2 /D8/CW/CT /BF/BH/B9/CS/CX/D1/CT/D2/D7/CX/D3/D2/CP/D0 /D7/D4/CP/CR/CT /CQ /CT/CR/CP/D9/D7/CT/D8/CW/CT /D4 /D3/D7/D8/CT/D6/CX/D3/D6 /CS/CX/D7/D8/D6/CX/CQ/D9/D8/CX/D3/D2 /CR/CP/D2/D2/D3/D8 /CQ /CT /D7/CW/CP/D6/D4 /CT/D6 /D8/CW/CP/D2 /D8/CW/CT /CX/D2/CS/CX/DA/CX/CS/D9/CP/D0 /D1/D3 /CS/CT/D0/D7 /CX/D2 /D8/CW/CT /D1/CX/DC/D8/D9/D6/CT /CP/D2/CS /D8/CW/CT/CX/D2/CS/CX/DA/CX/CS/D9/CP/D0 /D1/D3 /CS/CT/D0/D7 /D1 /D9/D7/D8 /CQ /CT /CQ/D6/D3/CP/CS/D0/DD /D8/D9/D2/CT/CS /D8/D3 /CP/D0/D0/D3 /DB/D8 /CW /CT /D1 /D8 /D3/CR /D3 /DA /CT/D6 /D8/CW/CT /BF/BH/B9/CS/CX/D1/CT/D2/D7/CX/D3/D2/CP/D0 /D7/D4/CP/CR/CT/BA

2212.04356.pdf

Robust Speech Recognition via Large-Scale Weak Supervision Alec Radford* 1Jong Wook Kim* 1Tao Xu1Greg Brockman1Christine McLeavey1Ilya Sutskever1 Abstract We study the capabilities of speech processing systems trained simply to predict large amounts of transcripts of audio on the internet. When scaled to 680,000 hours of multilingual and multitask supervision, the resulting models generalize well to standard benchmarks and are often competitive with prior fully supervised results but in a zero- shot transfer setting without the need for any fine- tuning. When compared to humans, the models approach their accuracy and robustness. We are releasing models and inference code to serve as a foundation for further work on robust speech processing. 1. Introduction Progress in speech recognition has been energized by the development of unsupervised pre-training techniques exem- plified by Wav2Vec 2.0 (Baevski et al., 2020). Since these methods learn directly from raw audio without the need for human labels, they can productively use large datasets of un- labeled speech and have been quickly scaled up to 1,000,000 hours of training data (Zhang et al., 2021), far more than the 1,000 or so hours typical of an academic supervised dataset. When fine-tuned on standard benchmarks, this approach has improved the state of the art, especially in a low-data setting. These pre-trained audio encoders learn high-quality repre- sentations of speech, but because they are purely unsuper- vised they lack an equivalently performant decoder mapping those representations to usable outputs, necessitating a fine- tuning stage in order to actually perform a task such as speech recognition1. This unfortunately limits their use- fulness and impact as fine-tuning can still be a complex process requiring a skilled practitioner. There is an addi- tional risk with requiring fine-tuning. Machine learning *Equal contribution1OpenAI, San Francisco, CA 94110, USA. Correspondence to: Alec Radford <alec@openai.com >, Jong Wook Kim <jongwook@openai.com >. 1Baevski et al. (2021) is an exciting exception - having devel- oped a fully unsupervised speech recognition systemmethods are exceedingly adept at finding patterns within a training dataset which boost performance on held-out data from the same dataset. However, some of these patterns are brittle and spurious and don’t generalize to other datasets and distributions. In a particularly disturbing example, Rad- ford et al. (2021) documented a 9.2% increase in object classification accuracy when fine-tuning a computer vision model on the ImageNet dataset (Russakovsky et al., 2015) without observing any improvement in average accuracy when classifying the same objects on seven other natural image datasets. A model that achieves “superhuman” per- formance when trained on a dataset can still make many basic errors when evaluated on another, possibly precisely because it is exploiting those dataset-specific quirks that humans are oblivious to (Geirhos et al., 2020). This suggests that while unsupervised pre-training has im- proved the quality of audio encoders dramatically, the lack of an equivalently high-quality pre-trained decoder, com- bined with a recommended protocol of dataset-specific fine- tuning, is a crucial weakness which limits their usefulness and robustness. The goal of a speech recognition system should be to work reliably “out of the box” in a broad range of environments without requiring supervised fine-tuning of a decoder for every deployment distribution. As demonstrated by Narayanan et al. (2018), Likhomanenko et al. (2020), and Chan et al. (2021) speech recognition sys- tems that are pre-trained in a supervised fashion across many datasets/domains exhibit higher robustness and generalize much more effectively to held-out datasets than models trained on a single source. These works achieve this by combining as many existing high-quality speech recogni- tion datasets as possible. However, there is still only a moderate amount of this data easily available. SpeechStew (Chan et al., 2021) mixes together 7 pre-existing datasets totalling 5,140 hours of supervision. While not insignifi- cant, this is still tiny compared to the previously mentioned 1,000,000 hours of unlabeled speech data utilized in Zhang et al. (2021). Recognizing the limiting size of existing high-quality super- vised datasets, recent efforts have created larger datasets for speech recognition. By relaxing the requirement of gold- standard human-validated transcripts, Chen et al. (2021) and Galvez et al. (2021) make use of sophisticated automatedarXiv:2212.04356v1 [eess.AS] 6 Dec 2022

Rombach-High-Resolution-Image-Synthesis-With-Latent-Diffusion-Models-CVPR-2022-paper.pdf

High-Resolution Image Synthesis with Latent Diffusion Models Robin Rombach1∗Andreas Blattmann1∗Dominik Lorenz1Patrick Esser Bj¨orn Ommer1 1Ludwig Maximilian University of Munich & IWR, Heidelberg University, Germany Runway ML https://github.com/CompVis/latent-diffusion Abstract By decomposing the image formation process into a se- quential application of denoising autoencoders, diffusion models (DMs) achieve state-of-the-art synthesis results on image data and beyond. Additionally, their formulation al- lows for a guiding mechanism to control the image gen- eration process without retraining. However, since these models typically operate directly in pixel space, optimiza- tion of powerful DMs often consumes hundreds of GPU days and inference is expensive due to sequential evalu- ations. To enable DM training on limited computational resources while retaining their quality and flexibility, we apply them in the latent space of powerful pretrained au- toencoders. In contrast to previous work, training diffusion models on such a representation allows for the first time to reach a near-optimal point between complexity reduc- tion and detail preservation, greatly boosting visual fidelity. By introducing cross-attention layers into the model archi- tecture, we turn diffusion models into powerful and flexi- ble generators for general conditioning inputs such as text or bounding boxes and high-resolution synthesis becomes possible in a convolutional manner. Our latent diffusion models (LDMs) achieve new state of the art scores for im- age inpainting and class-conditional image synthesis and highly competitive performance on various tasks, includ- ing unconditional image generation, text-to-image synthe- sis, and super-resolution, while significantly reducing com- putational requirements compared to pixel-based DMs. 1. Introduction Image synthesis is one of the computer vision fields with the most spectacular recent development, but also among those with the greatest computational demands. Espe- cially high-resolution synthesis of complex, natural scenes is presently dominated by scaling up likelihood-based mod- els, potentially containing billions of parameters in autore- gressive (AR) transformers [ 64,65]. In contrast, the promis- ing results of GANs [ 3,26,39] have been revealed to be mostly confined to data with comparably limited variability as their adversarial learning procedure does not easily scale to modeling complex, multi-modal distributions. Recently, diffusion models [ 79], which are built from a hierarchy of denoising autoencoders, have shown to achieve impressive ∗The first two authors contributed equally to this work.Inputours (f= 4) PSNR:27.4R-FID:0.58DALL-E ( f= 8) PSNR:22.8R-FID:32.01VQGAN ( f= 16 ) PSNR:19.9R-FID:4.98 Figure 1. Boosting the upper bound on achievable quality with less agressive downsampling. Since diffusion models offer excel- lent inductive biases for spatial data, we do not need the heavy spa- tial downsampling of related generative models in latent space, but can still greatly reduce the dimensionality of the data via suitable autoencoding models, see Sec. 3. Images are from the DIV2K [ 1] validation set, evaluated at 5122px. We denote the spatial down- sampling factor by f. Reconstruction FIDs [ 28] and PSNR are calculated on ImageNet-val. [ 12]; see also Tab. 8. results in image synthesis [ 29,82] and beyond [ 7,44,47,56], and define the state-of-the-art in class-conditional image synthesis [ 15,30] and super-resolution [ 70]. Moreover, even unconditional DMs can readily be applied to tasks such as inpainting and colorization [ 82] or stroke-based syn- thesis [ 52], in contrast to other types of generative mod- els [19,45,67]. Being likelihood-based models, they do not exhibit mode-collapse and training instabilities as GANs and, by heavily exploiting parameter sharing, they can model highly complex distributions of natural images with- out involving billions of parameters as in AR models [ 65]. Democratizing High-Resolution Image Synthesis DMs belong to the class of likelihood-based models, whose mode-covering behavior makes them prone to spend ex- cessive amounts of capacity (and thus compute resources) on modeling imperceptible details of the data [ 16,71]. Al- though the reweighted variational objective [ 29] aims to ad- dress this by undersampling the initial denoising steps, DMs are still computationally demanding, since training and evaluating such a model requires repeated function evalu- ations (and gradient computations) in the high-dimensional space of RGB images. As an example, training the most powerful DMs often takes hundreds of GPU days ( e.g. 150 - 1000 V100 days in [ 15]) and repeated evaluations on a noisy version of the input space render also inference expensive, 10684

2402.09668.pdf

How to Train Data-Efficient LLMs Noveen Sachdeva1 2Benjamin Coleman1Wang-Cheng Kang1Jianmo Ni1Lichan Hong1Ed H. Chi1 James Caverlee1 3Julian McAuley2Derek Zhiyuan Cheng1 Abstract The training of large language models (LLMs) is expensive. In this paper, we study data-efficient approaches for pre-training LLMs, i.e., techniques that aim to optimize the Pareto frontier of model quality and training resource/data consumption. We seek to understand the tradeoffs associated with data selection routines based on (i) expensive- to-compute data-quality estimates, and (ii) max- imization of coverage and diversity-based mea- sures in the feature space. Our first technique, ASK-LLM , leverages the zero-shot reasoning ca- pabilities of instruction-tuned LLMs to directly assess the quality of a training example. To tar- get coverage, we propose DENSITY sampling, which models the data distribution to select a diverse sample. In our comparison of 19sam- plers, involving hundreds of evaluation tasks and pre-training runs, we find that ASK-LLM and DENSITY are the best methods in their respec- tive categories. Coverage sampling can recover the performance of the full data, while models trained on ASK-LLM data consistently outper- form full-data training—even when we reject 90% of the original dataset, while converging up to 70% faster. 1. Introduction Large language model (LLM) pre-training is perhaps the most data- and compute-intensive task attempted by the machine learning community to date, with impressive capa- bilities primarily being accomplished by training massive transformer architectures on trillions of tokens of text (Ope- nAI, 2023; Gemini et al., 2023; Touvron et al., 2023b). But even these incredibly capable LLMs are subject to em- pirical scaling laws, which predict sharply diminishing re- turns from a linear increase in model- or data-size (Hoff- mann et al., 2022; Kaplan et al., 2020). Power-law scaling therefore acts as a soft limit on model quality, beyond which 1Google DeepMind2University of California, San Diego 3Texas A&M University. Correspondence to: Noveen Sachdeva <noveen@google.com>.it is prohibitively expensive to drive performance by scal- ing up the data or model. At the same time, Sorscher et al. (2022)—in the context of vision pre-training—show that we can significantly improve the power law constants in the aforementioned scaling laws if we prioritize important training examples using some robust notion of data quality or impact. A similar call for data-curation is also apparent in the context of training LLMs, where our largest models are quickly ap- proaching their capacity and data thresholds. LIMA (Zhou et al., 2023) showed that LLaMA-65B (Touvron et al., 2023a) can be better aligned with human preferences when trained on a set of 1,000 carefully selected fine-tuning prompts, compared to training on as much as 52,000 unfil- tered examples. Tirumala et al. (2023) recently conducted a large-scale data-efficient pre-training evaluation, showing that a 6.7B OPT model (Zhang et al., 2022) can converge up to 20% faster on data curated by a technique based on strati- fied cluster sampling. The Phi-2 experiments also suggest that when data curation is performed at a human-expert level (e.g., by textbook editors), models can outperform baselines that are up to 25x larger (Javaheripi et al., 2023). Data curation routines can be fundamentally characterized as selecting training samples for quality, coverage, or some mixture of both (Figure 2). In this work, we seek to under- stand how quality and coverage affect the data efficiency of LLM pre-training. Our core research question is: “Are cheap-to-compute heuristics like maximum- coverage enough to pre-train a SoTA LLM, or are there real benefits from costly samplers that carefully evaluate the quality of each example?” This question is crucial to answer because data-curation algorithms can improve the Pareto frontier of the data- quantity ↔model-quality tradeoff, directly addressing the bottleneck of power-law scaling by enabling higher-quality models to be trained using less data. Data curation also unlocks new tradeoffs between training time, inference cost, data collection effort, and downstream performance. For example, if we consider the compute-constrained (single- epoch) regime, a data-efficient LLM training routine may reach the desired performance using only X% of the data 1arXiv:2402.09668v1 [cs.LG] 15 Feb 2024

mapreduce.pdf

MapReduce: Simplied Data Processing onLargeClusters JeffreyDean andSanjay Ghema wat jeff@google.com, sanjay@google.com Google,Inc. Abstract MapReduce isaprogramming model andanassoci- ated implementation forprocessing andgenerating large data sets. Users specify amap function thatprocesses a key/valuepairtogenerate asetofintermediate key/value pairs, andareduce function thatmergesallintermediate values associated with thesame intermediate key.Many realworld tasks areexpressible inthismodel, asshown inthepaper . Programs written inthisfunctional style areautomati- cally parallelized andexecuted onalargecluster ofcom- modity machines. Therun-time system takescare ofthe details ofpartitioning theinput data, scheduling thepro- gram' sexecution across asetofmachines, handling ma- chine failures, andmanaging therequired inter-machine communication. This allowsprogrammers without any experience with parallel anddistrib uted systems toeas- ilyutilize theresources ofalargedistrib uted system. Our implementation ofMapReduce runs onalarge cluster ofcommodity machines andishighly scalable: atypical MapReduce computation processes manyter- abytes ofdata onthousands ofmachines. Programmers ndthesystem easy touse: hundreds ofMapReduce pro- grams havebeen implemented andupwards ofonethou- sand MapReduce jobs areexecuted onGoogle' sclusters everyday. 1Introduction Overthepast veyears, theauthors andmanyothers at Google haveimplemented hundreds ofspecial-purpose computations that process largeamounts ofrawdata, such ascrawled documents, web request logs, etc., to compute various kinds ofderiveddata, such asinverted indices, various representations ofthegraph structure ofweb documents, summaries ofthenumber ofpages crawled perhost, thesetofmost frequent queries inagivenday,etc. Most such computations areconceptu- allystraightforw ard. However,theinput data isusually largeandthecomputations havetobedistrib uted across hundreds orthousands ofmachines inorder tonish in areasonable amount oftime. Theissues ofhowtopar- allelize thecomputation, distrib utethedata, andhandle failures conspire toobscure theoriginal simple compu- tation with largeamounts ofcomple xcode todeal with these issues. Asareaction tothiscomple xity,wedesigned anew abstraction thatallowsustoexpress thesimple computa- tions wewere trying toperform buthides themessy de- tails ofparallelization, fault-tolerance, data distrib ution andload balancing inalibrary .Our abstraction isin- spired bythemap andreduce primiti vespresent inLisp andmanyother functional languages. Werealized that most ofourcomputations involvedapplying amap op- eration toeach logical “record” inourinput inorder to compute asetofintermediate key/value pairs, andthen applying areduce operation toallthevalues thatshared thesame key,inorder tocombine thederiveddata ap- propriately .Our useofafunctional model with user- specied map andreduce operations allowsustoparal- lelize largecomputations easily andtousere-execution astheprimary mechanism forfaulttolerance. Themajor contrib utions ofthisworkareasimple and powerful interf acethatenables automatic parallelization anddistrib ution oflarge-scale computations, combined with animplementation ofthisinterf acethat achie ves high performance onlargeclusters ofcommodity PCs. Section 2describes thebasic programming model and givesseveralexamples. Section 3describes animple- mentation oftheMapReduce interf acetailored towards ourcluster -based computing environment. Section 4de- scribes several renements oftheprogramming model thatwehavefound useful. Section 5hasperformance measurements ofourimplementation foravariety of tasks. Section 6explores theuseofMapReduce within Google including ourexperiences inusing itasthebasis Toappear inOSDI 2004 1

2311.00208.pdf

Transformers as Recognizers of Formal Languages: A Survey on Expressivity Lena Strobl Umeå University lena.strobl@umu.seWilliam Merrill New York University willm@nyu.eduGail Weiss EPFL gail.weiss@epfl.ch David Chiang University of Notre Dame dchiang@nd.eduDana Angluin Yale University dana.angluin@yale.edu Abstract As transformers have gained prominence in natural language processing, some re- searchers have investigated theoretically what problems they can and cannot solve, by treating problems as formal languages . Exploring questions such as this will help to compare transformers with other models, and transformer variants with one another, for various tasks. Work in this subarea has made considerable progress in recent years. Here, we undertake a comprehensive survey of this work, documenting the diverse as- sumptions that underlie different results and providing a unified framework for harmoniz- ing seemingly contradictory findings. 1 Introduction Transformers (Vaswani et al., 2017) have gained prominence in natural language processing (NLP), both in direct applications like machine transla- tion and in pretrained models like BERT (Devlin et al., 2019) and GPT (Radford et al., 2018; Brown et al., 2020; OpenAI, 2023). Consequently, some researchers have sought to investigate their theoreti- cal properties. Such studies can broadly be divided into studies of expressivity andtrainability . Studies of expressivity could be further divided into those from the perspectives of approximation theory and of formal language theory. The former (e.g., Yun et al., 2020) investigates transformers as approx- imators of various classes of functions , along the lines of the universal approximation theorem for feedforward neural networks (Hornik et al., 1989; Cybenko, 1989). The latter, which is the subject of this survey, investigates transformers as recog- nizers of formal languages – that is, the inputs are treated as sequences of discrete symbols, and crucially as sequences of unbounded length.The core research question in this subarea is: How can we characterize the expressivity of trans- formers in relation to various formal models, such as automata, boolean circuits or formal logic? Re- lated questions include: •How do transformers compare to other architec- tures, like recurrent neural networks (RNNs), in expressivity? •How do transformer variants compare to one another in expressivity? Some further questions, which are not addressed by the papers surveyed here but could be addressed by future work in this subarea, include: •What new transformer variants are suggested by formal models? •Do failure cases anticipated from formal models occur in practice? •What insights into the complexity of human lan- guage are offered by a characterization of trans- former expressivity? Interpreting theoretical transformer results is complex due to diverse assumptions. Many vari- ants of transformers exist in practice, and even more have been proposed in theory. Also, trans- formers can recognize or generate languages in various ways. These diverse assumptions lead to varied, even seemingly contradictory, results. This paper provides a comprehensive survey of theoretical results on the expressive power of trans- formers. Compared to the surveys of Ackerman and Cybenko (2020) and Merrill (2021, 2023), which cover convolutional neural nets (CNNs), RNNs, and transformers, this is a narrower, but deeper, survey on transformers only. It sets up a unified framework for talking about transformer variants (§4), reviews key topics related to formal languages (§6), and systematically surveys results in the literature, documenting their assumptions and claims (§7) and harmonizing seemingly con- tradictory findings. See Table 1 for a summary. 1arXiv:2311.00208v1 [cs.LG] 1 Nov 2023

2402.04833.pdf

Long Is More for Alignment: A Simple but Tough-to-Beat Baseline for Instruction Fine-Tuning Hao Zhao1Maksym Andriushchenko1Francesco Croce1Nicolas Flammarion1 Abstract There is a consensus that instruction fine-tuning of LLMs requires high-quality data, but what are they? LIMA (NeurIPS 2023) and AlpaGa- sus (ICLR 2024) are state-of-the-art methods for selecting such high-quality examples, either via manual curation or using GPT-3.5-Turbo as a quality scorer. We show that the extremely sim- ple baseline of selecting the 1,000 instructions with longest responses from standard datasets can consistently outperform these sophisticated meth- ods according to GPT-4 and PaLM-2 as judges, while remaining competitive on the Open LLM benchmarks that test factual knowledge. We demonstrate this for several state-of-the-art LLMs (Llama-2-7B, Llama-2-13B, and Mistral-7B) and datasets (Alpaca-52k and Evol-Instruct-70k). In addition, a lightweight refinement of such long instructions can further improve the abilities of the fine-tuned LLMs, and allows us to obtain the 2nd highest-ranked Llama-2-7B-based model on AlpacaEval 2.0 while training on only 1,000 ex- amples and no extra preference data. We also con- duct a thorough analysis of our models to ensure that their enhanced performance is not simply due to GPT-4’s preference for longer responses, thus ruling out any artificial improvement. In conclu- sion, our findings suggest that fine-tuning on the longest instructions should be the default baseline for any research on instruction fine-tuning. 1. Introduction Pre-trained large language models (LLMs) need to undergo an alignment phase (Askell et al., 2021; Bai et al., 2022a; Ouyang et al., 2022; Wang et al., 2022; Taori et al., 2023) to make them suitable for downstream tasks like user inter- action or question answering. While the details may vary, alignment often relies on supervised fine-tuning (SFT) on 1EPFL, Switzerland. Correspondence to: Hao Zhao <hao.zhao@epfl.ch >.a dataset of instruction-response pairs to improve conver- sational ability, followed by reinforcement learning from either human (RLHF) (Ouyang et al., 2022) or automated (RLAIF) (Bai et al., 2022b; Lee et al., 2023) feedback to pro- mote the preferred style and content of replies. It is an active research direction to study whether it is possible to achieve satisfactory results while relying only on SFT, which would avoid the (potentially expensive) process of collecting pref- erence data. Taori et al. (2023) created Alpaca, an open source dataset of 52k instruction-response pairs, and fine- tuned on it a Llama-2-7B model to match the performance of the closed-source text-davinci-003 model. Then, Chen et al. (2023) introduced AlpaGasus, consisting of the 9k examples of Alpaca which are judged of highest quality by GPT-3.5- Turbo, to further improve the instruction-following abilities of the fine-tuned models. The intuition that instruction fine- tuning (IFT) might benefit from fewer demonstrations but of higher quality has been further pursued by Zhou et al. (2023) which manually curated LIMA, a dataset of 1k examples, which outperforms AlpaGasus. While the quality of the instructions seems to play a major role for IFT, it remains unclear which are the distinguishing features of high quality demonstrations. In this work, we revisit the significant efforts in constructing instruction-tuning datasets from prior work. Inspired by the fact LIMA contains much longer examples than Alpaca and the observation of recent works (Singhal et al., 2023; Yuan et al., 2024) that RLHF and direct preference optimization (DPO) (Rafailov et al., 2023) seem to mostly make the out- puts longer, we test selecting longest responses as a simple and inexpensive heuristic to curate a small (only 1k exam- ples) and high-quality IFT dataset from a larger one. Sur- prisingly, fine-tuning a Llama-2-7B (Touvron et al., 2023) base model on the 1k longest elements of Alpaca outper- forms both AlpaGasus and LIMA in one-to-one comparison with different LLMs as judges and on the AlpacaEval 2.0 benchmark (see Fig. 1). Moreover, simply improving the quality and the style of the response in Alpaca-1k-longest with GPT-3.5-Turbo, in combination with NEFTune noise augmentation (Jain et al., 2023), allows us to obtain the the 2nd highest-ranked Llama-2-7B-based model on AlpacaE- val 2.0. In this case, our simple method yields models which surpass LLMs with the same base model but fine-tuned with 1arXiv:2402.04833v1 [cs.CL] 7 Feb 2024

1801.05134.pdf

Understanding the Disharmony between Dropout and Batch Normalization by Variance Shift Xiang Li1Shuo Chen1Xiaolin Hu2Jian Yang1 Abstract This paper first answers the question “why do the two most powerful techniques Dropout and Batch Normalization (BN) often lead to a worse performance when they are combined together?” in both theoretical and statistical aspects. The- oretically, we find that Dropout would shift the variance of a specific neural unit when we transfer the state of that network from train to test. How- ever, BN would maintain its statistical variance, which is accumulated from the entire learning procedure, in the test phase. The inconsistency of that variance (we name this scheme as “vari- ance shift”) causes the unstable numerical behav- ior in inference that leads to more erroneous pre- dictions finally, when applying Dropout before BN. Thorough experiments on DenseNet, ResNet, ResNeXt and Wide ResNet confirm our findings. According to the uncovered mechanism, we next explore several strategies that modifies Dropout and try to overcome the limitations of their com- bination by avoiding the variance shift risks. 1. Introduction (Srivastava et al., 2014) brought Dropout as a simple way to prevent neural networks from overfitting. It has been proved to be significantly effective over a large range of machine learning areas, such as image classification (Szegedy et al., 2015), speech recognition (Hannun et al., 2014) and even natural language processing (Kim et al., 2016). Before the birth of Batch Normalization, it became a necessity of almost all the state-of-the-art networks and successfully boosted their performances against overfitting risks, despite its amazing simplicity. (Ioffe & Szegedy, 2015) demonstrated Batch Normaliza- 1DeepInsight@PCALab, Nanjing University of Science and Technology, China2Tsinghua National Laboratory for Informa- tion Science and Technology (TNList) Department of Computer Science and Technology, Tsinghua University, China. Correspon- dence to: Xiang Li <xiang.li.implus@njust.edu.cn >. 𝑋=𝑥 ෠𝑋=𝑋−𝐸𝑀𝑜𝑣𝑖𝑛𝑔(𝑋) 𝑉𝑎𝑟𝑀𝑜𝑣𝑖𝑛𝑔𝑋+𝜀𝑋𝑉𝑎𝑟𝑇𝑟𝑎𝑖𝑛𝑋=1 𝑝 𝑉𝑎𝑟𝑇𝑒𝑠𝑡𝑋=1𝑉𝑎𝑟𝑀𝑜𝑣𝑖𝑛𝑔𝑋=𝐸(1 𝑝) 𝑉𝑎𝑟𝑀𝑜𝑣𝑖𝑛𝑔𝑋=𝐸(1 𝑝) 𝑥~𝒩(0,1)Train Mode Test Mode𝑋=𝑎1 𝑝𝑥𝑋 𝑥~𝒩(0,1)𝜇=𝐸𝑋,𝜎2=𝑉𝑎𝑟𝑋,෠𝑋=𝑋−𝜇 𝜎2+𝜀 𝐸𝑀𝑜𝑣𝑖𝑛𝑔𝑋←𝐸(𝜇)𝑉𝑎𝑟𝑀𝑜𝑣𝑖𝑛𝑔𝑋←𝐸(𝜎2)Dropout 𝑎~Bernoulli (𝑝) BN 0 20 40 60 80 100 BN layer index on DenseNet trained on CIFAR1000.51.01.52.02.53.03.5max(real_vari moving_vari,moving_vari real_vari)Test Acc 77.42%, No Dropout in each bottleneck Test Acc 68.55%, Dropout 0.5 in each bottleneckFigure 1. Up: a simplified mathematical illustration of “variance shift”. In test mode, the neural variance of Xis different from that in train mode caused by Dropout, yet BN attempts to regard that variance as the popular statistic accumulated from training. Note thatpdenotes for the Dropout retain ratio and acomes from Bernoulli distribution which has probability pof being 1. Down: variance shift in experimental statistics on DenseNet trained on CIFAR100 dataset. The curves are both calculated from the same training data . “moving vari” is the moving variance (take its mean value instead if it’s a vector) that the i-th BN layer accumulates dur- ing the entire learning, and “ real var i” stands for the real variance of neural response before the i-th BN layer in inference. tion (BN), a powerful skill that not only speeded up all the modern architectures but also improved upon their strong baselines by acting as regularizers. Therefore, BN has been implemented in nearly all the recent network structures (Szegedy et al., 2016; 2017; Howard et al., 2017; Zhang et al., 2017) and demonstrates its great practicability and effectiveness. However, the above two nuclear weapons always fail to obtain an extra reward when combined together practically. In fact, a network even performs worse and unsatisfactorily when it is equipped with BN and Dropout simultaneously. (Ioffe & Szegedy, 2015) have already realized that BN elim- inates the need for Dropout in some cases – the authors exposed the incompatibility between them, thus conjecturedarXiv:1801.05134v1 [cs.LG] 16 Jan 2018

2305.13301.pdf

TRAINING DIFFUSION MODELS WITH REINFORCEMENT LEARNING Kevin Black∗1Michael Janner∗1Yilun Du2Ilya Kostrikov1Sergey Levine1 1University of California, Berkeley2Massachusetts Institute of Technology {kvablack, janner, kostrikov, sergey.levine}@berkeley.edu yilundu@mit.edu ABSTRACT Diffusion models are a class of flexible generative models trained with an approximation to the log-likelihood objective. However, most use cases of diffusion models are not concerned with likelihoods, but instead with downstream objectives such as human-perceived image quality or drug effectiveness. In this paper, we investigate reinforcement learning methods for directly optimizing diffusion models for such objectives. We describe how posing denoising as a multi-step decision- making problem enables a class of policy gradient algorithms, which we refer to as denoising diffusion policy optimization ( DDPO ), that are more effective than alternative reward-weighted likelihood approaches. Empirically, DDPO can adapt text-to-image diffusion models to objectives that are difficult to express via prompting, such as image compressibility, and those derived from human feedback, such as aesthetic quality. Finally, we show that DDPO can improve prompt-image alignment using feedback from a vision-language model without the need for additional data collection or human annotation. The project’s website can be found athttp://rl-diffusion.github.io . 1 I NTRODUCTION Diffusion probabilistic models (Sohl-Dickstein et al., 2015) have recently emerged as the de facto standard for generative modeling in continuous domains. Their flexibility in representing complex, high-dimensional distributions has led to the adoption of diffusion models in applications including image and video synthesis (Ramesh et al., 2021; Saharia et al., 2022; Ho et al., 2022), drug and material design (Xu et al., 2021; Xie et al., 2021; Schneuing et al., 2022), and continuous control (Janner et al., 2022; Wang et al., 2022; Hansen-Estruch et al., 2023). The key idea behind diffusion models is to iteratively transform a simple prior distribution into a target distribution by applying a sequential denoising process. This procedure is conventionally motivated as a maximum likelihood estimation problem, with the objective derived as a variational lower bound on the log-likelihood of the training data. However, most use cases of diffusion models are not directly concerned with likelihoods, but instead with downstream objective such as human-perceived image quality or drug effectiveness. In this paper, we consider the problem of training diffusion models to satisfy such objectives directly, as opposed to matching a data distribution. This problem is challenging because exact likelihood computation with diffusion models is intractable, making it difficult to apply many conventional reinforcement learning (RL) algorithms. We instead propose to frame denoising as a multi-step decision-making task, using the exact likelihoods at each denoising step in place of the approximate likelihoods induced by a full denoising process. We present a policy gradient algorithm, which we refer to as denoising diffusion policy optimization ( DDPO ), that can optimize a diffusion model for downstream tasks using only a black-box reward function. We apply our algorithm to the finetuning of large text-to-image diffusion models. Our initial evaluation focuses on tasks that are difficult to specify via prompting, such as image compressibility, and those derived from human feedback, such as aesthetic quality. However, because many reward functions of interest are difficult to specify programmatically, finetuning procedures often rely on large-scale human labeling efforts to obtain a reward signal (Ouyang et al., 2022). In the case of text-to-image diffusion, we propose a method for replacing such labeling with feedback from a vision-language model (VLM). Similar to RLAIF finetuning for language models (Bai et al., 2022b), the resulting procedure allows for diffusion models to be adapted to reward functions that would otherwise require 1arXiv:2305.13301v3 [cs.LG] 1 Oct 2023

2306.04488.pdf

Rewarded soups: towards Pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards Alexandre Rame1∗, Guillaume Couairon1,2†, Mustafa Shukor1†, Corentin Dancette1†,Jean-Baptiste Gaya1,2†,Laure Soulier1,Matthieu Cord1,3 1Sorbonne Université, CNRS, ISIR, Paris, France2Meta AI3Valeo.ai Abstract Foundation models are first pre-trained on vast unsupervised datasets and then fine-tuned on labeled data. Reinforcement learning, notably from human feedback (RLHF), can further align the network with the intended usage. Yet the imperfec- tions in the proxy reward may hinder the training and lead to suboptimal results ; the diversity of objectives in real-world tasks and human opinions exacerbate the issue. This paper proposes embracing the heterogeneity of diverse rewards by following a multi-policy strategy. Rather than focusing on a single a priori reward, we aim for Pareto-optimal generalization across the entire space of preferences. To this end, we propose rewarded soup , first specializing multiple networks independently (one for each proxy reward) and then interpolating their weights linearly. This succeeds empirically because we show that the weights remain linearly connected when fine-tuned on diverse rewards from a shared pre-trained initialization. We demon- strate the effectiveness of our approach for text-to-text (summarization, Q&A, helpful assistant, review), text-image (image captioning, text-to-image generation, visual grounding, VQA), and control (locomotion) tasks. We hope to enhance the alignment of deep models, and how they interact with the world in all its diversity. 1 Introduction Foundation models [ 1] have emerged as the standard paradigm to learn neural networks’ weights. They are typically first pre-trained through self-supervision [ 2,3,4,5] and then fine-tuned [ 6,7] via supervised learning [ 8]. Yet, collecting labels is expensive, and thus supervision may not cover all possibilities and fail to perfectly align [ 9,10,11] the trained network with the intended applications. Recent works [ 12,13,14] showed that deep reinforcement learning (DRL) helps by learning from various types of rewards. A prominent example is reinforcement learning from human feedback (RLHF) [ 12,15,16,17], which appears as the current go-to strategy to refine large language models (LLMs) into powerful conversational agents such as ChatGPT [ 13,18]. After pre-training on next token prediction [ 19] using Web data, the LLMs are fine-tuned to follow instructions [ 20,21,22] before reward maximization. This RL strategy enhances alignment by evaluating the entire generated sentence instead of each token independently, handling the diversity of correct answers and allowing for negative feedback [ 23]. Similar strategies have been useful in computer vision (CV) [ 14,24], for instance to integrate human aesthetics into image generation [25, 26, 27]. Diversity of proxy rewards. RL is usually seen as more challenging than supervised training [ 28], notably because the real reward—ideally reflecting the users’ preferences—is often not specified at training time. Proxy rewards are therefore developed to guide the learning, either as hand-engineered metrics [ 29,30,31] or more recently in RLHF as models trained to reflect human preferences ∗Project lead, main contributor, correspondence to alexandre.rame@isir.upmc.fr. †Equal experimental contribution, order determined at random. Further information and resources related to this project can be found on this website. 37th Conference on Neural Information Processing Systems (NeurIPS 2023).arXiv:2306.04488v2 [cs.LG] 16 Oct 2023

2210.03057.pdf

LANGUAGE MODELS ARE MULTILINGUAL CHAIN -OF-THOUGHT REASONERS Freda Shi1,2,∗Mirac Suzgun1,3,∗Markus Freitag1Xuezhi Wang1 Suraj Srivats4Soroush Vosoughi4Hyung Won Chung1Yi Tay1 Sebastian Ruder1Denny Zhou1Dipanjan Das1Jason Wei1 1Google Research2Toyota Technological Institute at Chicago 3Stanford University4Dartmouth College ABSTRACT We evaluate the reasoning abilities of large language models in multilingual settings. We introduce the Multilingual Grade School Math (MGSM) bench- mark, by manually translating 250 grade-school math problems from the GSM8K dataset (Cobbe et al., 2021) into tentypologically diverse languages. We find that the ability to solve MGSM problems via chain-of-thought prompting emerges with increasing model scale, and that models have strikingly strong multilin- gual reasoning abilities, even in underrepresented languages such as Bengali and Swahili. Finally, we show that the multilingual reasoning abilities of lan- guage models extend to other tasks such as commonsense reasoning and word- in-context semantic judgment. The MGSM benchmark is publicly available at https://github.com/google-research/url-nlp . 0.01% 1% 100%010203040506070 Underrepresented languages (SW,BN,TE,TH)High-resource languages (JA,ZH,RU,ES,FR,DE)English (EN) Frequency of language in pre-training dataset (token percentage) MGSM Accuracy (%)Translate to English with Google Translate and solve with English intermediate steps Intermediate reasoning steps in the language of the question Intermediate reasoning steps in English Figure 1: Correlation between language frequency and MGSM accuracy for PaLM-540B. The accuracy is surprisingly high, even for underrepresented languages like Swahili ( SW) and Bengali (BN), which account for less than 0.01% of the pre-training dataset. ∗Equal contribution. Work done during internship at Google Research. 1arXiv:2210.03057v1 [cs.CL] 6 Oct 2022

2306.17806.pdf

Stay on topic with Classifier-Free Guidance Guillaume V . Sanchez* Hexaglobe EleutherAI gsanchez@hexaglobe.comHonglu Fan* University of Geneva EleutherAI honglu.fan@unige.chAlexander Spangher* Information Sciences Institute University of Southern California spangher@usc.edu Elad Levi Sightful eladlevico@gmail.comPawan Sasanka Ammanamanchi IIIT Hyderabad Eleuther AI pawansasanka@gmail.comStella Biderman Booz Allen Hamilton EleutherAI stellabiderman@gmail.com Abstract Classifier-Free Guidance (CFG) [ 37] has recently emerged in text-to-image generation as a lightweight technique to encourage prompt-adherence in generations. In this work, we demonstrate that CFG can be used broadly as an inference-time technique in pure language modeling. We show that CFG (1) improves the performance of Pythia, GPT-2 and LLaMA-family models across an array of tasks: Q&A, reasoning, code generation, and machine translation, achieving SOTA on LAMBADA with LLaMA-7B over PaLM-540B; (2) brings improvements equivalent to a model with twice the parameter-count; (3) can stack alongside other inference-time methods like Chain-of-Thought and Self-Consistency, yielding further improvements in difficult tasks; (4) can be used to increase the faithfulness and coherence of assistants in challenging form-driven and content-driven prompts: in a human evaluation we show a 75% preference for GPT4All using CFG over baseline. 1 Introduction “Today in France , citizens were celebrating Christmas” “Today in France , and chickens lay eggs” γ =0 γ =1 γ =1.5 “Today in France , citizens were celebrating Thanksgiving” x0x1“Today in France , citizens were celebrating Bastille Day” γ =0.5 Figure 1: A notional 2D projection of a textual latent space showing how increasing the guidance weight γincreases the importance of the prompt “Today in France,”.In recent years large language models have exhibited strong generative capabilities to solve a diverse range of tasks [ 26,15,71]. “Prompting” is typically used to con- dition generation, with task instructions and context [ 64], or a small set of examples [ 15]. However, language gener- ation, especially with smaller models, has been shown to struggle with issues such as hallucination [ 49], degrada- tion [ 38] and meandering [ 76]. Various approaches have been proposed to address this, e.g.: instruction-finetuning [81,70] and reinforcement learning [ 56,4,6]. These tech- niques are expensive and their compute and data cost may not be accessible to all users. In this paper we propose an inference time methodology which, as shown in Figure 1, gives more importance to the user intent, expressed through the prompt. Our hypothesis in this paper is: fo- cusing more on the prompt at inference-time will result in generations that better align with expected behavior. Text-to-image-generation, too, has been shown to suffer from similar problems [ 28]. Standard inference approaches can ignore parts of the prompt-conditioning, especially with specific or uncommon prompts [ 53]. Classifier Guidance [ 28] *These authors contributed equally to this workarXiv:2306.17806v1 [cs.CL] 30 Jun 2023

2310.10638v5.pdf

Published as a conference paper at ICLR 2024 IN-CONTEXT PRETRAINING : LANGUAGE MODELING BEYOND DOCUMENT BOUNDARIES Weijia Shi1,2Sewon Min1,2Maria Lomeli1Chunting Zhou1 Margaret Li1,2Gergely Szilvasy1Rich James1Xi Victoria Lin1 Noah A. Smith2,3Luke Zettlemoyer1,2Scott Yih1Mike Lewis1 1Meta AI2University of Washington3Allen Institute for AI swj0419@cs.washington.edu ABSTRACT Large language models (LMs) are currently trained to predict tokens given doc- ument prefixes, enabling them to directly perform long-form generation and prompting-style tasks which can be reduced to document completion. Existing pretraining pipelines train LMs by concatenating random sets of short documents to create input contexts but the prior documents provide no signal for predicting the next document. We instead present IN-CONTEXT PRETRAINING , a new approach where language models are pretrained on a sequence of related documents, thereby explicitly encouraging them to read and reason across document boundaries. We can do IN-CONTEXT PRETRAINING by simply changing the document ordering so that each context contains related documents, and directly applying existing pretraining pipelines. However, this document sorting problem is challenging. There are billions of documents and we would like the sort to maximize contextual similarity for every document without repeating any data. To do this, we intro- duce approximate algorithms for finding related documents with efficient nearest neighbor search and constructing coherent input contexts with a graph traversal algorithm. Our experiments show IN-CONTEXT PRETRAINING offers a simple and scalable approach to significantly enhance LMs’ performance: we see notable improvements in tasks that require more complex contextual reasoning, including in-context learning (+8%), reading comprehension (+15%), faithfulness to previous contexts (+16%), long-context reasoning (+5%), and retrieval augmentation (+9%). 1 I NTRODUCTION Large language models (LMs) are trained to complete documents; each token is predicted given the context provided by the prefix of the document it appears in. Such contexts can be widely varied, especially at pretraining scale, allowing models to excel on diverse tasks such as instruction- following (Ouyang et al., 2022), conversational interfaces (OpenAI, 2023), reading comprehen- sion (Zhang et al., 2020), and in-context learning (Brown et al., 2020). However, recent studies highlight that LMs sometimes struggle to understand more complex contexts: they can fail to follow instructions accurately (McKenzie et al., 2023; Efrat & Levy, 2020; Liu & Liu, 2023), struggle with reasoning over conditioned documents (Liu et al., 2023; Shi et al., 2023a), and exhibit high variance in in-context learning (Zhao et al., 2021). In this paper, we present IN-CONTEXT PRETRAINING , a new pretraining method that learns to predict tokens conditioned on a sequence of related documents, explicitly enabling the model to read and reason about much more varied and longer contexts that go beyond document boundaries. Current LM training pipelines concatenate random sets of shorter documents to create longer con- text windows. However, the prior documents provide no signal for predicting the next document, incurring unnecessary computational overhead for tokens that do not require communication between them (de Vries, 2023). IN-CONTEXT PRETRAINING instead reorders the pretraining data by combin- ing several semantically related documents to create a coherent input context, thereby exposing LMs to long relevant contexts and providing pretraining signals beyond document boundaries. We illustrate this via an example in Figure 1: when predicting the following tokens for the phrase “ For 2022, FIFA set the prize money at $42m, ” a previous document stating that the “ World Cup never awarded 1arXiv:2310.10638v5 [cs.CL] 9 Mar 2024

2305.15348.pdf

READ: Recurrent Adaptation of Large Transformers Sid Wang John Nguyen Ke Li Carole-Jean Wu Meta AI {yuwang2020,ngjhn,kli26,carolejeanwu}@meta.com Abstract Fine-tuning large-scale Transformers has led to the explosion of many AI applica- tions across Natural Language Processing and Computer Vision tasks. However, fine-tuning all pre-trained model parameters becomes impractical as the model size and number of tasks increase. Parameter-efficient transfer learning (PETL) methods aim to address these challenges. While effective in reducing the number of trainable parameters, PETL methods still require significant energy and compu- tational resources to fine-tune. In this paper, we introduce REcurrent ADaption (READ) — a lightweight and memory-efficient fine-tuning method — to overcome the limitations of the current PETL approaches. Specifically, READ inserts a small RNN network alongside the backbone model so that the model does not have to back-propagate through the large backbone network. Through comprehensive empirical evaluation of the GLUE benchmark, we demonstrate READ can achieve a56% reduction in the training memory consumption and an 84% reduction in the GPU energy usage while retraining high model quality compared to full-tuning. Additionally, the model size of READ does not grow with the backbone model size, making it a highly scalable solution for fine-tuning large Transformers. 1 Introduction READ Adapter LoRA BitFit Prompt Full-tuning0.00.20.40.60.81.0Normalized Energy Consumption (lower is better) Figure 1: The normalized energy consumption rel- ative to full-tuning on GLUE tasks.Large-scale transformers architecture have achieved state-of-the-art results in several Nat- ural Language Processing (NLP) tasks [ 2,5,22, 23,25,33]. Scaling up the size of these models has been shown to confer various benefits, such as improved model prediction performance and sample efficiency [9, 14, 34]. The conventional paradigm is to pre-train large-scale models on generic web-scale data and fine-tune the models to downstream tasks. However, fine-tuning these models has become prohibitively expensive. Since 2018, the model size has increased by almost two orders of magnitude faster than GPU memory [ 20], resulting in prohibitively high cost to advance AI technologies [ 36]. Only a few well-funded institutions have the resources to fine-tune these models. Parameter-efficient transfer learning (PETL) [ 1,13,15,16,18,19,38] has emerged as a promising solution to overcome the challenges of full fine-tuning. Parameter-efficient transfer learning techniques aim to address these challenges by leveraging smaller and more task-specific models to efficiently adapt the pre-trained model’s parameters to the target task. Additive (e.g., adapters): Inserting small modules into the transformer blocks [ 13]. Soft Prompts (e.g., prefix-tuning) [ 18,19]: Small parameters concatenated Preprint. Under review.arXiv:2305.15348v1 [cs.LG] 24 May 2023

2309.10668.pdf

Language Modeling Is Compression Grégoire Delétang*1, Anian Ruoss*1, Paul-Ambroise Duquenne2, Elliot Catt1, Tim Genewein1, Christopher Mattern1, Jordi Grau-Moya1, Li Kevin Wenliang1, Matthew Aitchison1, Laurent Orseau1, Marcus Hutter1and Joel Veness1 *Equal contributions,1Google DeepMind,2Meta AI & Inria It has long been established that predictive models can be transformed into lossless compressors and vice versa. Incidentally, in recent years, the machine learning community has focused on training increasingly large and powerful self-supervised (language) models. Since these large language models exhibit impressive predictive capabilities, they are well-positioned to be strong compressors. In this work, we advocate for viewing the prediction problem through the lens of compression and evaluate the compression capabilities of large (foundation) models. We show that large language models are powerful general-purpose predictors and that the compression viewpoint provides novel insights into scaling laws, tokenization, and in-context learning. For example, Chinchilla 70B, while trained primarily on text, compresses ImageNet patches to 43.4% and LibriSpeech samples to 16.4% of their raw size, beating domain-specific compressors like PNG (58.5%) or FLAC (30.3%), respectively. Finally, we show that the prediction-compression equivalence allows us to use any compressor (like gzip) to build a conditional generative model. 1. Introduction Information theory and machine learning are inextricably linked and have even been referred to as “two sides of the same coin” (MacKay, 2003). One particularly elegant connection is the essential equivalence between probabilistic models of data and lossless compression. The source coding theorem (Shannon, 1948) is the fundamental theorem describing this idea, i.e., the expected message length in bits of an optimal entropy encoder is equal to the negative log2-likelihood of the statistical model. In other words, maximizing the log2-likelihood (of the data) is equivalent to minimizing the number of bits required per message. Indeed, lossless compression with a probabilistic model can be achieved in a variety of different ways, including Huffman coding (Huffman, 1952), arithmetic coding (Pasco, 1977; Rissanen, 1976), and asymmetric numeral systems (Duda, 2009). Arithmetic coding, in particular, is known to be optimal in terms of coding length, meaning that the overall compression performance depends on the capabilities of the probabilistic model (Fig. 1). Incidentally,inrecentyears,largepre-trainedTransformers(Vaswanietal.,2017),so-called foundation models(Bommasani et al., 2021), have proven to be highly successful across a wide range of predictive tasks (Bubeck et al., 2023; Rae et al., 2021) and are thus promising candidates for use with arithmetic coding. Indeed, Transformer-based compression with arithmetic coding has produced state-of-the- art results both in the online (Bellard, 2021; Mao et al., 2022) and offline settings (Valmeekam et al., 2023). In the online setting, a pseudo-randomly initialized model is directly trained on the stream of data that is to be compressed, while the offline setting, which we consider in our work, trains the model on an external dataset before employing it to compress a (potentially different) data stream. Consequently, offline compression is performed in-context , with a fixed set of model parameters. Transformers have demonstrated impressive in-context learning abilities (Brown et al., 2020; Genewein et al., 2023; Laskin et al., 2023; Wei et al., 2022), which renders them ideally suited for offline compression. However, as we will discuss in this work, Transformers are actually trained to compress well, and therefore musthave good in-context learning abilities. Corresponding authors: {gdelt, anianr}@google.comarXiv:2309.10668v1 [cs.LG] 19 Sep 2023

2404.16710v1.pdf

LayerSkip: Enabling Early Exit Inference and Self-Speculative Decoding Mostafa Elhoushi1,†,∗,Akshat Shrivastava1,†,∗,Diana Liskovich2,†,Bram Wasti2,Basil Hosmer1, Liangzhen Lai3,Anas Mahmoud4,Bilge Acun1,Saurabh Agrawal6,Ahmed Roman7,Ahmed A Aly3,Beidi Chen1,5,Carole Jean-Wu1 1FAIR at Meta,2GenAI at Meta,3Reality Labs at Meta,4University of Toronto,5Carnegie Mellon University,6University of Wisconsin-Madison,7Dana-Farber Cancer Institute ∗Equal Contribution ,†Core Contributor We present LayerSkip, an end-to-end solution to speed-up inference of large language models (LLMs). First, during training we apply layer dropout, with low dropout rates for earlier layers and higher dropout rates for later layers, and an early exit loss where all transformer layers share the same exit. Second, during inference, we show that this training recipe increases the accuracy of early exit at earlier layers, without adding any auxiliary layers or modules to the model. Third, we present a novel self-speculative decoding solution where we exit at early layers and verify and correct with remaining layers of the model. Our proposed self-speculative decoding approach has less memory footprint than other speculative decoding approaches and benefits from shared compute and activations of the draft and verification stages. We run experiments on different Llama model sizes on different types of training: pretraining from scratch, continual pretraining, finetuning on specific data domain, and finetuning on specific task. We implement our inference solution and show speedups of up to 2.16 × on summarization for CNN/DM documents, 1.82 ×on coding, and 2.0 ×on TOPv2 semantic parsing task. Date:April 26, 2024 Correspondence: Mostafa Elhoushi, Akshat Shrivastava atmelhoushi@meta.com ,akshats@meta.com Code:In progress 1 Introduction Large Language Models (LLMs) have been deployed to many applications, yet their high compute and memory requirements lead to high financial and energy costs when deployed to GPU servers Samsi et al. (2023). Acceleration solutions do exist to deploy to commodity GPUs on laptops but they suffer from significant drop in accuracy Zhu et al. (2023). Accelerating LLMs further to mobile or edge devices is still an active research area Çöplü et al. (2023); Liu et al. (2024). While a large portion of LLM acceleration approaches reduce number of non-zero weights Xia et al. (2023) (a.k.a. sparsity), number of bits per weight Xiao et al. (2023) (a.k.a. quantization), number of heads per layer Shim et al. (2021) (a.k.a. head pruning), a smaller portion of approaches focus on reducing number of layers Fan et al. (2020); Elbayad et al. (2020). In this paper, we explore reducing the number of layers required for each token by exiting early during inference. Unlike quantization or sparsity, acceleration by reducing number of layers does not require specialized hardware or software kernels. Moreover, a popular research trend in LLM acceleration is speculative decoding Leviathan et al. (2023); Chen et al. (2023) that has no drop in accuracy, where a large model, referred to as the mainmodel, is accompanied with a faster model, referred to as the draftmodel. The advantage of speculative decoding is that it leads to faster inference compared to the main model, but requires a larger memory footprint and complexity in implementation to maintain key-value (KV) cache in two different models. In addition to exiting early, this paper also proposes combining exiting early with speculative decoding to propose a self-speculative decoding approach that does not require an additional model or auxiliary layers. 1arXiv:2404.16710v1 [cs.CL] 25 Apr 2024

2212.14024.pdf

DEMONSTRATE –SEARCH –PREDICT : Composing retrieval and language models for knowledge-intensive NLP Omar Khattab1Keshav Santhanam1Xiang Lisa Li1David Hall1 Percy Liang1Christopher Potts1Matei Zaharia1 Abstract Retrieval-augmented in-context learning has emerged as a powerful approach for addressing knowledge-intensive tasks using frozen language models (LM) and retrieval models (RM). Exist- ing work has combined these in simple “retrieve- then-read” pipelines in which the RM retrieves passages that are inserted into the LM prompt. To begin to fully realize the potential of frozen LMs and RMs, we propose DEMONSTRATE – SEARCH –PREDICT (DSP ), a framework that re- lies on passing natural language texts in sophisti- cated pipelines between an LM and an RM. DSP can express high-level programs that bootstrap pipeline-aware demonstrations, search for rele- vant passages, and generate grounded predictions, systematically breaking down problems into small transformations that the LM and RM can handle more reliably. We have written novel DSP pro- grams for answering questions in open-domain, multi-hop, and conversational settings, establish- ing in early evaluations new state-of-the-art in- context learning results and delivering 37–120%, 8–39%, and 80–290% relative gains against the vanilla LM (GPT-3.5), a standard retrieve-then- read pipeline, and a contemporaneous self-ask pipeline, respectively. We release DSP athttps: //github.com/stanfordnlp/dsp . 1. Introduction In-context learning adapts a frozen language model (LM) to tasks by conditioning the LM on a textual prompt including task instructions and a few demonstrating examples (Mc- Cann et al., 2018; Radford et al., 2019; Brown et al., 2020). For knowledge-intensive tasks such as question answering, fact checking, and information-seeking dialogue, retrieval models (RM) are increasingly used to augment prompts 1Stanford University . Correspondence to: Omar Khattab <okhattab@cs.stanford.edu >. Preprint . How many storeys are in the castle David Gregory inherited? LM:Castle Gregory has three storeys.❌Hallucinates a fictitious castle RM: “St. Gregory Hotel is a nine-floor boutique hotel in D.C...” LM: St. Gregory Hotel has nine storeys.❌Retrieves a different building LM: “Which castle did David Gregory inherit?” RM: “David Gregory inherited Kinnairdy Castle in 1664...” LM: “How many storyes does Kinnairdy Castle have?” RM: “Kinnairdy Castle is a tower house, having five storeys…” LM: Kinnairdy Castle has fivestoreys.Vanilla LM Retrieve- then-Read Multi-Hop DSP ProgramFigure 1. A comparison between three systems based on GPT- 3.5 (text-davinci-002 ). On its own, the LM often makes false assertions. An increasingly popular retrieve-then-read pipeline fails when simple search can’t find an answer. In contrast, a task- aware DSP program successfully decomposes the problem and produces a correct response. Texts edited for presentation. with relevant information from a large corpus (Lazaridou et al., 2022; Press et al., 2022; Khot et al., 2022). Recent work has shown such retrieval-augmented in-context learning to be effective in simple “retrieve-then-read” pipelines: a query is fed to the RM and the retrieved pas- sages become part of a prompt that provides context for the LM to use in its response. In this work, we argue that the fact that both LMs and RMs consume (and generate or retrieve) natural language texts creates an opportunity for much more sophisticated interactions between them. Fully realizing this would be transformative: frozen LMs and RMs could serve as infrastructure across tasks, enabling ML- and domain-experts alike to rapidly build grounded AI systems at a high level of abstraction and with lower deployment overheads and annotation costs. Figure 1 begins to illustrate the power of retrieval- augmented in-context learning, but also the limitations of “retrieve-then-read” (Lazaridou et al., 2022; Izacard et al., 2022). Our query is “How many storeys are in the castle David Gregory inherited?” When prompted to answer this, GPT-3.5 ( text-davinci-002 ; Ouyang et al. 2022) makes up a fictitious castle with incorrect attributes, highlighting the common observation that knowledge stored in LM pa- rameters is often unreliable (Shuster et al., 2021; Ishii et al., 2022). Introducing an RM component helps, as the LM can ground its responses in retrieved passages, but a rigidarXiv:2212.14024v2 [cs.CL] 23 Jan 2023

L08_expressivity.pdf

Expressive Variational Autoencoders John Thickstun The Gaussian VAE parameterizes the prior r(z), conditional likelihood p(x|z), and posterior approximation q(x|z) with with Gaussian distributions. The in-expressivity of these Gaussian models can make it difficult to capture the distribution p(x); complaints about the “blurriness” of the VAE may be attributable to these assumptions. Note that many papers visualize the mean gθ(˜z) of the decoder network, rather than samples gθ(˜z) +η, which coupled with a Gaussian noise model onXcould exacerbate blurriness. PixelCNN and PixelVAE One way to increase the expressivity of the VAE is to remove the conditional-independence as- sumption from the decoder distribution p(x|z). In the standard Gaussian VAE, the components xi ofxare conditionally independent given the latent code z: p(x|z) =|X|∏ i=1p(xi|z) =|X|∏ i=1N(xi|µi(z),σ2). (1) We can remove this assumption by building a fully-autoregressive model of the decoder distribution over observations x, i.e. p(x|z) =|X|∏ i=1p(xi|x<i,z). (2) An auto-regressive parameterization of the conditional likelihood called PixelVAE is explored by Gulrajani et al. [2017], based on a line of work building autoregressive models called PixelCNN [van den Oord et al., 2016b,a, Salimans et al., 2017] that extends the NADE modeling perspective to images. One oddity of these models is that, in order to construct an autoregressive factorization of the like distribution over images, we need to fix a (somewhat arbitrary) ordering over pixels; the standard choice is to order the pixels from left to right, top-to-bottom, starting with the pixel in the upper-left corner of the image. One might question whether the order matters; while any order leads to a valid factorization of the joint distribution, perhaps some factorizations would be easier to learn than others? This question was asked in the original NADE work, and the answer. There is followup work on orderless NADE [Uria et al., 2014] that learns an ensemble of factored autoregressive models, one for each possible ordering of pixels; by ensembling these models, it may be possible to construct a better model than using any particular ordering. But in practice, just picking an arbitrary ordering doesn’t seem to cause too much trouble. Two serious problems with using autoregressive likelihoods p(x|z) are posterior collapse (dis- cussed in the next section) and the computational expense of sampling from an autoregressive 1

2311.11944v1.pdf

FINANCE BENCH : A New Benchmark for Financial Question Answering Pranab Islam1∗Anand Kannappan1Douwe Kiela2,3 Rebecca Qian1Nino Scherrer1Bertie Vidgen1 1Patronus AI2Contextual AI3Stanford University Abstract FINANCE BENCH is a first-of-its-kind test suite for evaluating the performance of LLMs on open book financial question answering (QA). It comprises 10,231 questions about publicly traded companies, with corresponding an- swers and evidence strings. The questions inFINANCE BENCH are ecologically valid and cover a diverse set of scenarios. They are in- tended to be clear-cut and straightforward to answer to serve as a minimum performance standard. We test 16 state of the art model con- figurations (including GPT-4-Turbo, Llama2 and Claude2, with vector stores and long con- text prompts) on a sample of 150 cases from FINANCE BENCH , and manually review their answers (n=2,400). The cases are available open-source. We show that existing LLMs have clear limitations for financial QA. Notably, GPT-4-Turbo used with a retrieval system in- correctly answered or refused to answer 81% of questions. While augmentation techniques such as using longer context window to feed in relevant evidence improve performance, they are unrealistic for enterprise settings due to in- creased latency and cannot support larger fi- nancial documents. We find that all models examined exhibit weaknesses, such as halluci- nations, that limit their suitability for use by enterprises. 1 Introduction Finance specialists routinely need to find informa- tion about companies and industries, summarize and analyze that information, and then reason about it. This time-intensive and difficult work is cru- cial for making investment decisions, developing financial strategies, and conducting due diligence. Large Language Models (LLMs) have the poten- tial to augment and automate labor-intensive parts of financial analysis because of their impressive capabilities in natural language understanding, rea- soning, and writing (Nori et al., 2023; Bubeck et al., ∗Authors are ordered alphabetically Figure 1: Incorrect model responses (using a shared vector store) to a question in FINANCE BENCH . The correct answer is given by the human expert. 2023). However, a key challenge blocking the fi- nancial industry’s adoption of LLMs is that there are few ways of evaluating models’ performance on finance‘-specific tasks. And, without rigorous, systematic, and measurable evaluation processes, the industry cannot (1) understand the strengths and weaknesses of models; (2) assess whether they perform well enough to use in high-stakes live set- tings; and (3) track how their capabilities change over time. The financial domain presents unique challenges for LLMs. First, models need domain-specific knowledge about financial topics and terminology, as well as companies and industries. It is unclear how much financial information and statistics ap- pear in the pre-training data of models. In part to address models’ lack of knowledge about finance, BloombergGPT was released in March 2023 as the first LLM specialised for the financial domain (Wu et al., 2023). Second, models need up-to-date financial information and to understand relevant financial news. However, many models’ data isarXiv:2311.11944v1 [cs.CL] 20 Nov 2023

2403.09636.pdf

Dynamic Memory Compression: Retrofitting LLMs for Accelerated Inference Piotr Nawrot*Q VAdrian Ła ´ncucki*Q KMarcin ChochowskiQDavid TarjanQEdoardo M. PontiV QNVIDIAKUniversity of WrocławVUniversity of Edinburgh Abstract Transformers have emerged as the backbone of large language models (LLMs). However, genera- tion remains inefficient due to the need to store in memory a cache of key–value representations for past tokens, whose size scales linearly with the input sequence length and batch size. As a solu- tion, we propose Dynamic Memory Compression (DMC), a method for on-line key–value cache compression at inference time. Most importantly, the model learns to apply different compression rates in different heads and layers. We retrofit pre- trained LLMs such as Llama 2 (7B, 13B and 70B) into DMC Transformers, achieving up to ~3.7 × throughput increase during auto-regressive infer- ence on an NVIDIA H100 GPU. DMC is applied via continued pre-training on a negligible percent- age of the original data without adding any extra parameters. We find that DMC preserves the origi- nal downstream performance with up to 4 ×cache compression, outperforming up-trained grouped- query attention (GQA). GQA and DMC can be even combined to obtain compounded gains. As a result DMC fits longer contexts and larger batches within any given memory budget. 1. Introduction Transformer Large Language Models (LLMs) are the state of the art in generative and conversational AI (Touvron et al., 2023; Jiang et al., 2023). Their deployment, however, is curtailed in part by their inefficiency. This is not only due to the quadratic complexity of attention layers (Bahdanau et al., 2014; Vaswani et al., 2017): during generation, Trans- formers store the keys and values of past tokens in memory to avoid recomputing them multiple times. Since this key– value (KV) cache grows linearly with the sequence length and batch size, generation with Transformers quickly be- *Equal contribution. Correspondence to: Piotr Nawrot < piotr.nawrot@ed.ac.uk >. Key-value cachekvt(append)(a) Regular key–value cache with items kvidepicted as boxes. New items are always appended. kvtαt=1 (accumulate) weighted averageαt=0 (append) Key-value cache kvt (b) Dynamic Memory Compression (DMC) chooses whether to accumulate or append current items, resulting in a smaller key– value cache. Figure 1: Key–value cache update mechanisms. comes prohibitive due to the excessive memory load. This issue emerges even more clearly with long-context genera- tion (e.g., in dialogues and stories) or when serving large numbers of user queries. A widespread solution to increase the memory efficiency of Transformers during inference is Grouped Query Attention (GQA; Ainslie et al., 2023; Shazeer, 2019), which uses a number of key and value heads inferior to the number of query heads through parameter sharing. As an alternative, the number of overall tokens in memory can be reduced through token merging (Zhang et al., 2018; Liu et al., 2018; Bolya et al., 2022) or token pruning (Anagnostidis et al., 2023; Kim & Cho, 2020). Nevertheless, these methods often pay the price of a severe degradation in downstream performance. On the other hand, hardware/IO-aware (Dao et al., 2022; Kwon et al., 2023) and sub-quadratic algorithms for attention (Beltagy et al., 2020; Choromanski et al., 2020) do not alleviate the memory load of the KV cache. In our work, we aim to achieve a lossless compression of the KV cache of LLMs, thus retaining their performance while reducing their memory load. To this end, we propose Dy- namic Memory Compression (DMC). As shown in Figure 1, during every time step, DMC decides whether to append 1arXiv:2403.09636v1 [cs.CL] 14 Mar 2024

1610.03518v1.pdf

Transfer from Simulation to Real World through Learning Deep Inverse Dynamics Model Paul Christiano, Zain Shah, Igor Mordatch, Jonas Schneider, Trevor Blackwell, Joshua Tobin, Pieter Abbeel, and Wojciech Zaremba OpenAI, San Francisco, CA, USA Abstract — Developing control policies in simulation is often more practical and safer than directly running experiments in the real world. This applies to policies obtained from planning and optimization, and even more so to policies obtained from reinforcement learning, which is often very data demanding. However, a policy that succeeds in simulation often doesnt work when deployed on a real robot. Nevertheless, often the overall gist of what the policy does in simulation remains valid in the real world. In this paper we investigate such settings, where the sequence of states traversed in simulation remains reasonable for the real world, even if the details of the controls are not, as could be the case when the key differences lie in detailed friction, contact, mass and geometry properties. During execution, at each time step our approach computes what the simulation-based control policy would do, but then, rather than executing these controls on the real robot, our approach computes what the simulation expects the resulting next state(s) will be, and then relies on a learned deep inverse dynamics model to decide which real-world action is most suitable to achieve those next states. Deep models are only as good as their training data, and we also propose an approach for data collec- tion to (incrementally) learn the deep inverse dynamics model. Our experiments shows our approach compares favorably with various baselines that have been developed for dealing with simulation to real world model discrepancy, including output error control and Gaussian dynamics adaptation. I. I NTRODUCTION Many methods exist for generating control policies in simulated environments, including methods based on motion planning, optimization, control, and learning. However, an important practical challenge is that often there are discrep- ancies between simulation and the real world, which results in policies that work well in simulation yet perform poorly in the real world. Significant bodies of work exist that strive to address this challenge. One important line of work studies how to improve simulators to better match reality, which involves improving simulation of contact, non-rigidity, friction, as well as improving identification of physical quantities needed for accurate simulation such as mass, geometry, friction coefficients, elasticity. However, despite significant progress, discrepancies continue to exist, and more accurate simulation can have the downside of being slower. Another important line of work studies robustness of con- trol policies, which could be measured through, for example, gain and phase margins, and robust control methods exist that can optimize for these. Optimizing for robustness means finding control policies that apply across a wide range ofpossible real worlds, but unfortunately tends to come at the expense of performance in the one specific real world the system is faced with. Adaptive methods, which is the topic of this paper, do not use the same policy for the entire family of possible environments, but rather try to learn about the specific real world the system is faced with. In principle, such methods can exploit the physics of the real world and behave in the optimal way. Concretely, our work considers the following problem setting: We assume to be given a simulator and a method for generating policies that perform well in simulation. The goal is to leverage this to perform well in new real-world situations. To achieve this, a training period exists during which an adaptation mechanism can be trained to learn to adapt from simulation to real world by collecting experience on the real system, but without having access to the new real-world situations that the system will be evaluated on later. We leverage the following intuition: Often policies found from simulation capture the high-level gist well (e.g., overall trajectory), but fail to accurately capture some of the lower- level details, such as friction, stiction, backlash, hysteresis, precise measurements, precise deformation, etc. Indeed, this is the type of situation that motivates the work in this paper and in which we will be evaluating our approach (as well as baselines). Note that while we assume that a method exists for generating policies in simulation, our approach is agnostic to the details of this method, which could be based on any techniques from motion planning, optimization, control, learning, and others, which return a policy, which could be a model-predictive policy which uses the simulator in its inner loop. Our approach proceeds as follows: During execution on a test trajectory, at each time step it computes what the simulation-based control policy would do, but then, rather than executing these controls on the real robot, our approach computes what the simulation expects the next state(s) will be, and then relies on a learned deep inverse dynamics model to decide which real-world action is most suitable to achieve those next states. As our experiments show, when these inverse dynamics models are trained on sufficient data, this results in compelling transfer from simulation to real world, in particular with challenging dynamics involvingarXiv:1610.03518v1 [cs.RO] 11 Oct 2016

2302.03764.pdf

Sketchy: Memory-efficient Adaptive Regularization with Frequent Directions Vladimir Feinberg1Xinyi Chen1 2Y. Jennifer Sun2Rohan Anil1Elad Hazan1 2 Abstract Adaptive regularization methods that exploit more than the diagonal entries exhibit state of the art performance for many tasks, but can be pro- hibitive in terms of memory and running time. We find the spectra of the Kronecker-factored gra- dient covariance matrix in deep learning (DL) training tasks are concentrated on a small lead- ing eigenspace that changes throughout training, motivating a low-rank sketching approach. We describe a generic method for reducing memory and compute requirements of maintaining a ma- trix preconditioner using the Frequent Directions (FD) sketch. Our technique allows interpolation between resource requirements and the degrada- tion in regret guarantees with rank k: in the online convex optimization (OCO) setting over dimen- siond, we match full-matrix d2memory regret using onlydkmemory up to additive error in the bottomd−keigenvalues of the gradient covari- ance. Further, we show extensions of our work to Shampoo, placing the method on the memory- quality Pareto frontier of several large scale bench- marks. 1. Introduction DL optimization commonly relies on adaptive gradient methods, namely the Adam optimizer (Kingma & Ba, 2015). It differs from stochastic gradient descent in that the learn- ing rate is a structured diagonal matrix built from previous gradients rather than a scalar. In full matrix AdaGrad (Duchi et al., 2011) the inverse matrix square root of the sum of outer products of previous gradients is the learning rate. Full matrix preconditioning is impractical for modern deep learning architectures: for instance, the ResNet-50 archi- tecture (He et al., 2016) has over 23 million parameters, requiring more than 2 petabytes to represent its gradient covariance. Thus, diagonal preconditioning methods remain 1Google Research, Brain Team2Princeton University. Corre- spondence to: Vladimir Feinberg <vladf@google.com >. Preliminary work.popular. However, previous work has demonstrated state- of-the-art results in some settings, such as large-batch data parallel training, for nondiagonal forms of preconditioning (Martens & Grosse, 2015; Gupta et al., 2018; Agarwal et al., 2019; Chen et al., 2019; Anil et al., 2019; 2020). Further- more, as hardware evolves, memory efficiency becomes an increasing concern, as “logic improves much faster than wires and SRAM, so logic is relatively free” (Jouppi et al., 2021): from TPUv2 to TPUv3, per-chip bfloat16 oper- ations per second improved 2.67×but memory bandwidth only improved 1.29×. GPUs exhibit a similar pattern for compute and memory increase, at 5×and2.2×, for V100 to A100 (Dally et al., 2021). 0 20 40 60 80 100 training completion (%)020406080100mass (%)eigenvalue mass in in top 256 of 1024 eigs architecture conformer gnn resnet side left right Figure 1: Low-rank nuclear norm relative error. We tune ResNet-50, a Conformer, and a Graph Neural Net (GNN), with Shampoo across three different datasets (see Sec. 5.1). For a 2D layer with gradients G, Shampoo tracks the expo- nential moving average of factors GG⊤andG⊤G(left and right sides). We select a 1024×1024 covariance factor C across all these architectures for both sides and plot the pro- portion of spectral mass captured by the top 256eigenvalues, i.e.,∑256 i=1λi(C)/∑1024 i=1λi(C). Spectral investigation into the Kronecker-factored gradi- ent covariance matrix reveals a concentrated, but changing, spectrum (Fig. 1), suggesting the majority of the spectral mass can be represented by a low-rank matrix, albeit rotating over time. The Frequent Directions (FD) sketch provides a mechanism to track the top eigenvectors without materi- alizing the full covariance matrix (Ghashami et al., 2016).arXiv:2302.03764v1 [stat.ML] 7 Feb 2023

1608.04471.pdf

Stein Variational Gradient Descent: A General Purpose Bayesian Inference Algorithm Qiang Liu Dilin Wang Department of Computer Science Dartmouth College Hanover, NH 03755 {qiang.liu, dilin.wang.gr}@dartmouth.edu Abstract We propose a general purpose variational inference algorithm that forms a natural counterpart of gradient descent for optimization. Our method iteratively trans- ports a set of particles to match the target distribution, by applying a form of functional gradient descent that minimizes the KL divergence. Empirical studies are performed on various real world models and datasets, on which our method is competitive with existing state-of-the-art methods. The derivation of our method is based on a new theoretical result that connects the derivative of KL divergence under smooth transforms with Stein’s identity and a recently proposed kernelized Stein discrepancy, which is of independent interest. 1 Introduction Bayesian inference provides a powerful tool for modeling complex data and reasoning under uncer- tainty, but casts a long standing challenge on computing intractable posterior distributions. Markov chain Monte Carlo (MCMC) has been widely used to draw approximate posterior samples, but is often slow and has difficulty accessing the convergence. Variational inference instead frames the Bayesian inference problem into a deterministic optimization that approximates the target distribution with a simpler distribution by minimizing their KL divergence. This makes variational methods efficiently solvable by using off-the-shelf optimization techniques, and easily applicable to large datasets (i.e., "big data") using the stochastic gradient descent trick [e.g., 1]. In contrast, it is much more challenging to scale up MCMC to big data settings [see e.g., 2, 3]. Meanwhile, both the accuracy and computational cost of variational inference critically depend on the set of distributions in which the approximation is defined. Simple approximation sets, such as these used in the traditional mean field methods, are too restrictive to resemble the true posterior distributions, while more advanced choices cast more difficulties on the subsequent optimization tasks. For this reason, efficient variational methods often need to be derived on a model-by-model basis, causing is a major barrier for developing general purpose, user-friendly variational tools applicable for different kinds of models, and accessible to non-ML experts in application domains. This case is in contrast with the maximum a posteriori (MAP) optimization tasks for finding the posterior mode (sometimes known as the poor man’s Bayesian estimator , in contrast with the full Bayesian inference for approximating the full posterior distribution), for which variants of (stochastic) gradient descent serve as a simple, generic, yet extremely powerful toolbox. There has been a recent growth of interest in creating user-friendly variational inference tools [e.g., 4–7], but more efforts are still needed to develop more efficient general purpose algorithms. In this work, we propose a new general purpose variational inference algorithm which can be treated as a natural counterpart of gradient descent for full Bayesian inference (see Algorithm 1). Our algorithm uses a set of particles for approximation, on which a form of (functional) gradient descentarXiv:1608.04471v3 [stat.ML] 9 Sep 2019

1812.11118.pdf

Reconciling modern machine learning practice and the bias-variance trade-off Mikhail Belkina, Daniel Hsub, Siyuan Maa, and Soumik Mandala aThe Ohio State University, Columbus, OH bColumbia University, New York, NY September 12, 2019 Abstract Breakthroughs in machine learning are rapidly changing science and society, yet our fun- damental understanding of this technology has lagged far behind. Indeed, one of the central tenets of the field, the bias-variance trade-off, appears to be at odds with the observed behavior of methods used in the modern machine learning practice. The bias-variance trade-off implies that a model should balance under-fitting and over-fitting: rich enough to express underlying structure in data, simple enough to avoid fitting spurious patterns. However, in the modern practice, very rich models such as neural networks are trained to exactly fit (i.e., interpolate) the data. Classically, such models would be considered over-fit, and yet they often obtain high accuracy on test data. This apparent contradiction has raised questions about the mathematical foundations of machine learning and their relevance to practitioners. In this paper, we reconcile the classical understanding and the modern practice within a unified performance curve. This “double descent” curve subsumes the textbook U-shaped bias- variance trade-off curve by showing how increasing model capacity beyond the point of inter- polation results in improved performance. We provide evidence for the existence and ubiquity of double descent for a wide spectrum of models and datasets, and we posit a mechanism for its emergence. This connection between the performance and the structure of machine learning models delineates the limits of classical analyses, and has implications for both the theory and practice of machine learning. E-mail: mbelkin@cse.ohio-state.edu , djhsu@cs.columbia.edu , masi@cse.ohio-state.edu , mandal.32@osu.edu 1arXiv:1812.11118v2 [stat.ML] 10 Sep 2019

2002.05616.pdf

Learning the Stein Discrepancy for Training and Evaluating Energy-Based Models without Sampling Will Grathwohl1Kuan-Chieh Wang1J¨orn-Henrik Jacobsen1David Duvenaud1Richard Zemel1 Abstract We present a new method for evaluating and train- ing unnormalized density models. Our approach only requires access to the gradient of the unnor- malized model’s log-density. We estimate the Stein discrepancy between the data density p(x) and the model density q(x)defined by a vector function of the data. We parameterize this func- tion with a neural network and fit its parameters to maximize the discrepancy. This yields a novel goodness-of-fit test which outperforms existing methods on high dimensional data. Furthermore, optimizingq(x)to minimize this discrepancy pro- duces a novel method for training unnormalized models which scales more gracefully than exist- ing methods. The ability to both learn and com- pare models is a unique feature of the proposed method. 1. Introduction Energy-Based Models (EBMs), also known as unnormal- ized density models, are perhaps the most flexible way to parameterize a density. They hinge on the observation that any densityp(x)can be expressed as p(x) =exp(−E(x)) Z, (1) whereE:RD→R, known as the energy function , maps each point to a scalar, and Z=∫ xexp(−E(x))is the normalizing constant. A major benefit of EBMs is that they allow maximal free- dom in designing the energy function E. This makes it straightforward to incorporate prior knowledge about the problem, such as symmetries or domain-specific de- sign choices, into the structure of the model. This has 1University of Toronto and Vector Institute, Toronto, Canada. Correspondence to: Will Grathwohl <wgrath- wohl@cs.toronto.edu >. Proceedings of the 37thInternational Conference on Machine Learning , Vienna, Austria, PMLR 119, 2020. Copyright 2020 by the author(s). Figure 1. Density models trained with approximate MCMC sam- plers can fail to match the data density while still generating high- quality samples. Samples from approximate MCMC samplers follow a different distribution than the density they are applied to. It is this induced distribution which is trained to match the data. In contrast, our approach LSD directly matches the model density to the data density without reliance on a sampler. made EBMs an appealing candidate for applications in physics (No ´e et al., 2019), biology (Ingraham et al., 2019), neuroscience (Scellier & Bengio, 2017), and computer vi- sion (LeCun et al., 2007; Osadchy et al., 2007; Xie et al., 2016; 2019; 2018), to name a few. Despite their many benefits, EBMs present a central chal- lenge which complicates their use: because we cannot effi- ciently compute the normalizing constant, we cannot com- pute likelihoods under our model, making training and eval- uation difficult. Much prior work on EBMs has relied on MCMC sampling techniques to estimate the likelihood (for evaluation) and its gradient (for training). Other approaches train EBMs by finding easier-to-compute surrogate objec- tives which have similar optima to the maximum likelihood objective. These include Score Matching (Hyv ¨arinen, 2005) and Noise-Contrastive Estimation (Gutmann & Hyv ¨arinen, 2010). These original sampling- and score-based approaches were not able to scale to large, high-dimensional datasets as well as subsequently developed alternative models, such as Vari- ational Autoencoders (V AEs) (Kingma & Welling, 2013) and Normalizing Flows (NFs) (Rezende & Mohamed, 2015). These approaches offer more easily scalable training, evalua- tion, and sampling, but do so at the cost of a more restrictive model parameterization which can lead to well-known prob-arXiv:2002.05616v4 [stat.ML] 14 Aug 2020

2304.14802.pdf

ResiDual: Transformer with Dual Residual Connections Shufang Xie‡†, Huishuai Zhang†, Junliang Guo†, Xu Tan†∗, Jiang Bian† Hany Hassan Awadalla†,Arul Menezes†,Tao Qin†,Rui Yan‡∗ †Microsoft Research†Microsoft Azure Translation ‡Gaoling School of Artificial Intelligence, Renmin University of China {shufangxie,ruiyan}@ruc.edu.cn , {huzhang,junliangguo,xuta,jiabia,hanyh,arulm,taoqin}@microsoft.com Abstract Transformer networks have become the preferred architecture for many tasks due to their state-of-the-art performance. However, the optimal way to implement residual connections in Transformer, which are essential for effective training, is still debated. Two widely used variants are the Post-Layer Normalization (Post-LN) and Pre-Layer Normalization (Pre-LN) Transformers, which apply layer normal- ization after each residual block’s output or before each residual block’s input, respectively. While both variants enjoy their advantages, they also suffer from severe limitations: Post-LN causes gradient vanishing issue that hinders training deep Transformers, and Pre-LN causes representation collapse issue that limits model capacity. In this paper, we propose ResiDual, a novel Transformer archi- tecture with Pre-Post-LN (PPLN), which fuses the connections in Post-LN and Pre-LN together, and inherits their advantages while avoids their limitations. We conduct both theoretical analyses and empirical experiments to verify the effec- tiveness of ResiDual. Theoretically, we prove that ResiDual has a lower bound on the gradient to avoid the vanishing issue due to the residual connection from Pre-LN. Moreover, ResiDual also has diverse model representations to avoid the collapse issue due to the residual connection from Post-LN. Empirically, ResiDual outperforms both Post-LN and Pre-LN on several machine translation benchmarks across different network depths and data sizes. Thanks to the good theoretical and empirical performance, ResiDual Transformer can serve as a foundation architec- ture for different AI models (e.g., large language models). Our code is available at https://github.com/microsoft/ResiDual . 1 Introduction Transformer (Vaswani et al., 2017) has emerged as a powerful neural network architecture that has been successfully applied in various AI tasks, including machine translation (Vaswani et al., 2017), language model ing and generation (Radford et al., 2018, 2019; Brown et al., 2020), image recognition (Dosovitskiy et al., 2020), and speech synthesis (Ren et al., 2019). Despite its success, researchers are still exploring ways to further enhance its performance and deepen the understanding of its inner workings (Wang et al., 2019; Katharopoulos et al., 2020; Fedus et al., 2021). Among them, one area of ongoing research is the study of residual connections in the Transformer architecture (Liu et al., 2020; Xiong et al., 2020; Bachlechner et al., 2021). Two variants of residual connections have been proposed since the introduction of the Transformer, known as Post-LN and Pre-LN. The Post-LN variant applies layer normalization (LN) operations after the output of each residual block. ∗Corresponding Authors: Xu Tan, xuta@microsoft.com ; Rui Yan, ruiyan@ruc.edu.cn . Preprint. Under review.arXiv:2304.14802v1 [cs.CL] 28 Apr 2023

2403.07816.pdf

Branch-Train-MiX: Mixing Expert LLMs into a Mixture-of-Experts LLM Sainbayar Sukhbaatar ,Olga Golovneva ,Vasu Sharma ,Hu Xu,Xi Victoria Lin ,Baptiste Rozière ,Jacob Kahn,Daniel Li,Wen-tau Yih ,Jason Weston ,Xian Li FAIR at Meta We investigate efficient methods for training Large Language Models (LLMs) to possess capabilities in multiple specialized domains, such as coding, math reasoning and world knowledge. Our method, named Branch-Train-MiX (BTX), starts from a seed model, which is branched to train experts in embarrassingly parallel fashion with high throughput and reduced communication cost. After individual experts are asynchronously trained, BTX brings together their feedforward parameters as experts in Mixture-of-Expert (MoE) layers and averages the remaining parameters, followed by an MoE-finetuning stage to learn token-level routing. BTX generalizes two special cases, the Branch-Train-Merge method, which does not have the MoE finetuning stage to learn routing, and sparse upcycling, which omits the stage of training experts asynchronously. Compared to alternative approaches, BTX achieves the best accuracy-efficiency tradeoff. Date:March 13, 2024 Correspondence: {sainbar,xianl}@meta.com 1 Introduction In recent years, Large Language Models (LLMs) have shown impressive performance in a wide-range of tasks (Brown et al., 2020; Touvron et al., 2023; Achiam et al., 2023), including code generation (Li et al., 2022b; Rozière et al., 2023), solving math problems (Azerbayev et al., 2023), multilinguality (Zhao et al., 2024), etc. Training such LLMs requires a large amount of compute and data, exceeding thousands of GPUs and trillions of tokens. The training parallelization is typically done by maintaining multiple copies of the model on different GPUs and keeping them synchronized after each weight update. The cost of this frequent communication is the main bottleneck in scaling the training to more GPUs. Besides this issue, synchronized training is more vulnerable to hardware failures as a single failed GPU can cause the whole training to halt (Zhang et al., 2022; Gemini Team, 2023). Recent work by Li et al. (2022a) proposed the Branch-Train-Merge (BTM) method for embarrassingly parallel training of LLMs without any synchronization for improving the throughput of pretraining. It starts by creating multiple copies of a seed LLM, then separately training each copy on different subsets of data. This results in multiple independent LLMs that do not share any parameters and each LLM is an expert specializing in its own data distribution, such as knowledge domains, languages or even modalities. At test time, an input prompt is classified into one or more of the domains, and then the final outputs are formed from the corresponding expert models which are combined to predict the next token. While this approach makes training more efficient, its main drawback is the lack of a unified single model making it impossible to do further supervised finetuning (SFT) or reinforcement learning from human feedback (RLHF) finetuning (Ouyang et al., 2022), both of which can boost performance further, and are crucial steps in building aligned LLMs. A separate line of work for reducing the computational footprint of LLMs is the Mixture-of-Experts (MoE) approach (Jacobs et al., 1991; Shazeer et al., 2017), where only a subset of parameteters are active at any given time. In particular, MoE is applied to the feedforward sublayer of Transformers (Fedus et al., 2022; Roller et al., 2021; Lewis et al., 2021), allowing the total number of parameters to grow without additional computation. LLMs scaled in this way have shown impressive performance on downstream tasks (Jiang et al., 2024; Xue et al., 2024). Unlike Branch-Train-Merge, Mixture-of-Experts are often trained in a fully 1arXiv:2403.07816v1 [cs.CL] 12 Mar 2024

2209.15634.pdf

A General Framework for Sample-Efficient Function Approximation in Reinforcement Learning Zixiang Chen‡∗Chris Junchi Li⋄∗Angela Yuan‡∗Quanquan Gu‡Michael I. Jordan⋄,† Department of Computer Sciences, University of California, Los Angeles‡ Department of Electrical Engineering and Computer Sciences, University of California, Berkeley⋄ Department of Statistics, University of California, Berkeley† October 3, 2022 Abstract With the increasing need for handling large state and action spaces, general function approximation has become a key technique in reinforcement learning (RL). In this paper, we propose a general framework that unifies model-based and model-free RL, and an Admissible Bellman Characterization (ABC) class that subsumes nearly all Markov Decision Process (MDP) models in the literature for tractable RL. We propose a novel estimation function with decomposable structural properties for optimization-based exploration and the functional eluder dimension as a complexity measure of the ABC class. Under our framework, a new sample-efficient algorithm namely OPtimization-based ExploRation with Approximation (OPERA) is proposed, achieving regret bounds that match or improve over the best-known results for a variety of MDP models. In particular, for MDPs with low Witness rank, under a slightly stronger assumption, OPERA improves the state-of-the-art sample complexity results by a factor of dH. Our framework provides a generic interface to design and analyze new RL models and algorithms. 1 Introduction Reinforcement learning (RL) is a decision-making process that seeks to maximize the expected reward when an agent interacts with the environment [Sutton and Barto, 2018]. Over the past decade, RL has gained increasing attention due to its successes in a wide range of domains, including Atari games [Mnih et al., 2013], Go game [Silver et al., 2016], autonomous driving [Yurtsever et al., 2020], Robotics [Kober et al., 2013], etc. Existing RL algorithms can be categorized into value-based algorithms such as Q-learning [Watkins, 1989] and policy-based algorithms such as policy gradient [Sutton et al., 1999]. They can also be categorized as a model-free approach where one directly models the value function classes, or alternatively, a model-based approach where one needs to estimate the transition probability. Due to the intractably large state and action spaces that are used to model the real-world complex environment, function approximation in RL has become prominent in both algorithm design and theoretical analysis. It is a pressing challenge to design sample-efficient RL algorithms with general function approximations. In the special case where the underlying Markov Decision Processes (MDPs) enjoy certain linear structures, several lines of works have achieved polynomial sample complexity and/or√ Tregret guarantees under either model-free or model-based RL settings. For linear MDPs where the transition probability and the reward function admit linear structure, Yang and Wang [2019] developed a variant ofQ-learning when granted access to a generative model, Jin et al. [2020] proposed an LSVI-UCB algorithm with a ˜O(√ d3H3T) regret bound and Zanette et al. [2020a] further extended the MDP model and improved the regret to ˜O(dH√ T). Another line of work considers linear mixture MDPs Yang and 1arXiv:2209.15634v1 [cs.LG] 30 Sep 2022

2205.13147.pdf

Matryoshka Representation Learning Aditya Kusupati∗†⋄, Gantavya Bhatt∗†, Aniket Rege∗†, Matthew Wallingford†, Aditya Sinha⋄, Vivek Ramanujan†, William Howard-Snyder†, Kaifeng Chen⋄, Sham Kakade‡, Prateek Jain⋄and Ali Farhadi† †University of Washington,⋄Google Research,‡Harvard University {kusupati,ali}@cs.washington.edu ,prajain@google.com Abstract Learned representations are a central component in modern ML systems, serv- ing a multitude of downstream tasks. When training such representations, it is often the case that computational and statistical constraints for each down- stream task are unknown. In this context, rigid fixed-capacity representations can be either over or under-accommodating to the task at hand. This leads us to ask: can we design a flexible representation that can adapt to multiple down- stream tasks with varying computational resources? Our main contribution is Matryoshka Representation Learning (MRL ) which encodes information at different granularities and allows a single embedding to adapt to the computational constraints of downstream tasks. MRL minimally modifies existing representation learning pipelines and imposes no additional cost during inference and deployment. MRL learns coarse-to-fine representations that are at least as accurate and rich as independently trained low-dimensional representations. The flexibility within the learned Matryoshka Representations offer: (a) up to 14×smaller embedding size for ImageNet-1K classification at the same level of accuracy; (b) up to 14× real-world speed-ups for large-scale retrieval on ImageNet-1K and 4K; and (c) up to2% accuracy improvements for long-tail few-shot classification, all while being as robust as the original representations. Finally, we show that MRL extends seam- lessly to web-scale datasets (ImageNet, JFT) across various modalities – vision (ViT, ResNet), vision + language (ALIGN) and language (BERT). MRL code and pretrained models are open-sourced at https://github.com/RAIVNLab/MRL . 1 Introduction Learned representations [ 57] are fundamental building blocks of real-world ML systems [ 66,91]. Trained once and frozen, d-dimensional representations encode rich information and can be used to perform multiple downstream tasks [ 4]. The deployment of deep representations has two steps: (1) an expensive yet constant-cost forward pass to compute the representation [ 29] and (2) utilization of the representation for downstream applications [ 50,89]. Compute costs for the latter part of the pipeline scale with the embedding dimensionality as well as the data size ( N) and label space ( L). At web-scale [ 15,85] this utilization cost overshadows the feature computation cost. The rigidity in these representations forces the use of high-dimensional embedding vectors across multiple tasks despite the varying resource and accuracy constraints that require flexibility. Human perception of the natural world has a naturally coarse-to-fine granularity [ 28,32]. However, perhaps due to the inductive bias of gradient-based training [ 84], deep learning models tend to diffuse “information” across the entire representation vector. The desired elasticity is usually enabled in the existing flat and fixed representations either through training multiple low-dimensional models [ 29], jointly optimizing sub-networks of varying capacity [ 9,100] or post-hoc compression [ 38,60]. Each of these techniques struggle to meet the requirements for adaptive large-scale deployment either ∗Equal contribution – AK led the project with extensive support from GB and AR for experimentation. 36th Conference on Neural Information Processing Systems (NeurIPS 2022).arXiv:2205.13147v4 [cs.LG] 8 Feb 2024

2307.15043.pdf

Universal and Transferable Adversarial Attacks on Aligned Language Models Andy Zou1,2, Zifan Wang2, Nicholas Carlini3, Milad Nasr3, J. Zico Kolter1,4, Matt Fredrikson1 1Carnegie Mellon University,2Center for AI Safety, 3Google DeepMind,4Bosch Center for AI Abstract Because “out-of-the-box” large language models are capable of generating a great deal of objectionable content, recent work has focused on aligning these models in an attempt to prevent undesirable generation. While there has been some success at cir- cumventing these measures—so-called “jailbreaks” against LLMs—these attacks have required significant human ingenuity and are brittle in practice. Attempts at automatic adversarial prompt generation have also achieved limited success. In this paper, we propose a simple and effective attack method that causes aligned language models to generate objectionable behaviors. Specifically, our approach finds a suffix that, when attached to a wide range of queries for an LLM to produce objectionable content, aims to maximize the probability that the model produces an affirmative response (rather than refusing to answer). However, instead of relying on manual engineering, our ap- proach automatically produces these adversarial suffixes by a combination of greedy and gradient-based search techniques, and also improves over past automatic prompt generation methods. Surprisingly, we find that the adversarial prompts generated by our approach are highly transferable , including to black-box, publicly released, production LLMs. Specif- ically, we train an adversarial attack suffix on multiple prompts (i.e., queries asking for many different types of objectionable content), as well as multiple models (in our case, Vicuna-7B and 13B). When doing so, the resulting attack suffix induces objec- tionable content in the public interfaces to ChatGPT, Bard, and Claude , as well as open source LLMs such as LLaMA-2-Chat, Pythia, Falcon, and others. Inter- estingly, the success rate of this attack transfer is much higher against the GPT-based models, potentially owing to the fact that Vicuna itself is trained on outputs from ChatGPT. In total, this work significantly advances the state-of-the-art in adversarial attacks against aligned language models, raising important questions about how such systems can be prevented from producing objectionable information. Code is available atgithub.com/llm-attacks/llm-attacks . 1arXiv:2307.15043v2 [cs.CL] 20 Dec 2023

2207.10551.pdf

Scaling Laws vs Model Architectures : How does Inductive Bias Influence Scaling? Yi Tay∗Mostafa Dehghani∗Samira Abnar Hyung Won Chung William Fedus Jinfeng Rao Sharan Narang Vinh Q. Tran Dani Yogatama†Donald Metzler Google Research & DeepMind† {yitay,dehghani}@google.com Abstract There have been a lot of interest in the scal- ing properties of Transformer models (Kaplan et al., 2020). However, not much has been done on the front of investigating the effect of scaling properties of different inductive bi- ases and model architectures. Do model ar- chitectures scale differently? If so, how does inductive bias affect scaling behaviour? How does this influence upstream (pretraining) and downstream (transfer)? This paper conducts a systematic study of scaling behaviour of ten diverse model architectures such as Transform- ers, Switch Transformers, Universal Trans- formers, Dynamic convolutions, Performers, and recently proposed MLP-Mixers. Via ex- tensive experiments, we show that (1) archi- tecture is an indeed an important considera- tion when performing scaling and (2) the best performing model can fluctuate at different scales. We believe that the findings outlined in this work has significant implications to how model architectures are currently evaluated in the community. 1 Introduction There have been a lot recent interest in the scaling properties of Transformer models (Kaplan et al., 2020; Hernandez et al., 2021; Bahri et al., 2021; Henighan et al., 2020; Tay et al., 2021b; Abnar et al., 2021). However, not much is understood about the scaling properties of different inductive biases imposed by model architectures. Improve- ments at a a specific scale (compute, size etc) are often assumed to transfer to different scales and compute regions (So et al., 2019; Choromanski et al., 2020; Lan et al., 2019; Dehghani et al., 2018) and new research is often presented in a point-wise fashion with respect to scale. In short, it is not un- common for new methods to be presented with data points at very specific or limited compute regions ∗Yi and Mostafa contributed equally. Samira is now at Apple.(e.g., base size). We believe that understanding the interaction between architecture and scaling laws is crucial as designing models that perform well at diverse scales will likely have significant impact. This paper is an attempt to understand the ef- fect of inductive bias (architecture) on scaling laws of language models. To this end, we pre-train and finetune over ten diverse model architectures across multiple compute region and scales (e.g., from 15M to 40 Billion parameters). In total, we pre-train and finetune over 100 different models of different ar- chitectures and sizes and present insights and chal- lenges at scaling these ten diverse architectures. We consider a broad spectrum of models in our extensive experiments. Concretely, we con- sider several well-established Transformer vari- ants (Vaswani et al., 2017) such as Evolved Trans- former (So et al., 2019), Universal Transformers (Dehghani et al., 2018) and Switch Transformers (Fedus et al., 2021). We also consider lightweight models such as ALBERT (Lan et al., 2019) and/or efficient Transformers (Tay et al., 2020) such as Performer (Choromanski et al., 2020) and Funnel Transformers (Dai et al., 2020). In our comparison, we are also interested in finding out if general im- provements to the Transformer architectures such as Mixture-of-Softmax (Yang et al., 2017) and/or Gated Linear Units (Dauphin et al., 2017; Shazeer, 2020) influence the scaling behaviour of models. Finally, we also evaluate models outside the fam- ily of Transformers including Lightweight convo- lutions (Wu et al., 2019), Dynamic convolutions (Wu et al., 2019) and the recently proposed MLP- Mixers (Tolstikhin et al., 2021). Figure 1 illustrates an overview about the experiments we run. We also note that scaling these models is not as straightforward as it seems, i.e., there are intricate details of scale that are intertwined with architec- tural choices which we study in detail in this pa- per. For example, a distinct feature of Universal Transformers (and ALBERT) is parameter sharing.arXiv:2207.10551v1 [cs.LG] 21 Jul 2022

2212.14024v2.pdf

DEMONSTRATE –SEARCH –PREDICT : Composing retrieval and language models for knowledge-intensive NLP Omar Khattab1Keshav Santhanam1Xiang Lisa Li1David Hall1 Percy Liang1Christopher Potts1Matei Zaharia1 Abstract Retrieval-augmented in-context learning has emerged as a powerful approach for addressing knowledge-intensive tasks using frozen language models (LM) and retrieval models (RM). Exist- ing work has combined these in simple “retrieve- then-read” pipelines in which the RM retrieves passages that are inserted into the LM prompt. To begin to fully realize the potential of frozen LMs and RMs, we propose DEMONSTRATE – SEARCH –PREDICT (DSP ), a framework that re- lies on passing natural language texts in sophisti- cated pipelines between an LM and an RM. DSP can express high-level programs that bootstrap pipeline-aware demonstrations, search for rele- vant passages, and generate grounded predictions, systematically breaking down problems into small transformations that the LM and RM can handle more reliably. We have written novel DSP pro- grams for answering questions in open-domain, multi-hop, and conversational settings, establish- ing in early evaluations new state-of-the-art in- context learning results and delivering 37–120%, 8–39%, and 80–290% relative gains against the vanilla LM (GPT-3.5), a standard retrieve-then- read pipeline, and a contemporaneous self-ask pipeline, respectively. We release DSP athttps: //github.com/stanfordnlp/dsp . 1. Introduction In-context learning adapts a frozen language model (LM) to tasks by conditioning the LM on a textual prompt including task instructions and a few demonstrating examples (Mc- Cann et al., 2018; Radford et al., 2019; Brown et al., 2020). For knowledge-intensive tasks such as question answering, fact checking, and information-seeking dialogue, retrieval models (RM) are increasingly used to augment prompts 1Stanford University . Correspondence to: Omar Khattab <okhattab@cs.stanford.edu >. Preprint . How many storeys are in the castle David Gregory inherited? LM:Castle Gregory has three storeys.❌Hallucinates a fictitious castle RM: “St. Gregory Hotel is a nine-floor boutique hotel in D.C...” LM: St. Gregory Hotel has nine storeys.❌Retrieves a different building LM: “Which castle did David Gregory inherit?” RM: “David Gregory inherited Kinnairdy Castle in 1664...” LM: “How many storyes does Kinnairdy Castle have?” RM: “Kinnairdy Castle is a tower house, having five storeys…” LM: Kinnairdy Castle has fivestoreys.Vanilla LM Retrieve- then-Read Multi-Hop DSP ProgramFigure 1. A comparison between three systems based on GPT- 3.5 (text-davinci-002 ). On its own, the LM often makes false assertions. An increasingly popular retrieve-then-read pipeline fails when simple search can’t find an answer. In contrast, a task- aware DSP program successfully decomposes the problem and produces a correct response. Texts edited for presentation. with relevant information from a large corpus (Lazaridou et al., 2022; Press et al., 2022; Khot et al., 2022). Recent work has shown such retrieval-augmented in-context learning to be effective in simple “retrieve-then-read” pipelines: a query is fed to the RM and the retrieved pas- sages become part of a prompt that provides context for the LM to use in its response. In this work, we argue that the fact that both LMs and RMs consume (and generate or retrieve) natural language texts creates an opportunity for much more sophisticated interactions between them. Fully realizing this would be transformative: frozen LMs and RMs could serve as infrastructure across tasks, enabling ML- and domain-experts alike to rapidly build grounded AI systems at a high level of abstraction and with lower deployment overheads and annotation costs. Figure 1 begins to illustrate the power of retrieval- augmented in-context learning, but also the limitations of “retrieve-then-read” (Lazaridou et al., 2022; Izacard et al., 2022). Our query is “How many storeys are in the castle David Gregory inherited?” When prompted to answer this, GPT-3.5 ( text-davinci-002 ; Ouyang et al. 2022) makes up a fictitious castle with incorrect attributes, highlighting the common observation that knowledge stored in LM pa- rameters is often unreliable (Shuster et al., 2021; Ishii et al., 2022). Introducing an RM component helps, as the LM can ground its responses in retrieved passages, but a rigidarXiv:2212.14024v2 [cs.CL] 23 Jan 2023

2302.12441.pdf

MUX-PLMs: Data Multiplexing for High-throughput Language Models Vishvak Murahari1Ameet Deshpande1Carlos E. Jimenez1 Izhak Shafran2Mingqiu Wang2Yuan Cao2Karthik Narasimhan1 1Princeton University2Google Brain murahari@cs.princeton.edu Abstract The widespread adoption of large language models such as ChatGPT and Bard has led to unprecedented demand for these technolo- gies. The burgeoning cost of inference for ever- increasing model sizes coupled with hardware shortages has limited affordable access and poses a pressing need for efficiency approaches geared towards high throughput and perfor- mance. Multi-input multi-output (MIMO) al- gorithms such as data multiplexing, offer a promising solution with a many-fold increase in throughput by performing inference for mul- tiple inputs at the cost of a single input. Yet these approaches are not currently performant enough to be deployed in modern systems. We change that by developing MUX-PLMs, a class of high throughput pre-trained language models (PLMs) trained with data multiplexing, that can be fine-tuned for any downstream task to yield high-throughput high-performance. Our novel multiplexing and demultiplexing modules profi- ciently entangle and disentangle inputs, and en- able high-performance high throughput MUX- PLMs that are competitive with vanilla PLMs while achieving 2x/5x inference speedup with only a 1−4%drop on a broad suite of tasks.1 1 Introduction Language models like ChatGPT (OpenAI, 2023), PaLM (Chowdhery et al., 2022), T5 (Raffel et al., 2020), and CM3 (Aghajanyan et al., 2022), have seen unprecedented adoption in diverse sectors ranging from education and healthcare to manu- facturing and marketing. The proficiency of these tools has led to unprecedented demand for these models, with users facing frequent outages and ca- pacity limits. Additionally, ever-increasing model sizes and hardware shortages have constrained models’ ability to handle a very high load of re- quests, thus limiting large-scale affordable access 1Code + Models: https://github .com/ princeton-nlp/datamux-pretraining/ .to these models. These trends bring into focus the need for high-throughput, high-performance, ef- ficient, and environmentally responsible models that can be deployed at scale to meet the quickly growing demand. Multi-input Multi-output architectures (MIMO) (Havasi et al., 2021; Ramé et al., 2021; Murahari et al., 2022) are a promising hardware-agnostic and architecture-agnostic paradigm that perform inference for multiple inputs simultaneously at the cost of a single input. This efficiency paradigm is natively geared towards yielding high-throughput models, in addition to being complementary in ap- proach and motivation to current efficiency meth- ods such as pruning, quantization, and distilla- tion. Interestingly, MIMO approaches are partly inspired by the human brain’s extraordinary abil- ity to process multiple inputs and propagate in- formation at a high bandwidth with a few neural codes (Blumhagen et al., 2011; Akam and Kull- mann, 2014; Pirschel and Kretzberg, 2016; Hong et al., 2016; Friedrich et al., 2004). Murahari et al. (2022) introduced data multiplex- ing, a MIMO technique that can enable a many-fold increase in throughput. The method compresses Ndifferent instances into a single “multiplexed” hidden representation before decompressing it into Nindependent predictions. While they show the plausibility of MIMO training, their method leads to a significant drop in performance ( 20−30% points) compared to state-of-the-art models. In this work, we introduce MUX-PLMs, a class of high-throughput pre-trained language models trained in a MIMO fashion with data multiplex- ing to process multiple inputs (2-10) simultane- ously with a forward pass over a single instance. MUX-PLMs offer up to 400% improvement in throughput over baseline pre-trained models while only being ∼4points and ∼2points worse than baseline pre-trained language models for text clas- sification and token classification tasks, respec-arXiv:2302.12441v2 [cs.LG] 22 May 2023

10.1038.s41467-021-25756-4.pdf

ARTICLE Efficient generative modeling of protein sequences using simple autoregressive models Jeanne Trinquier1,2, Guido Uguzzoni3,4, Andrea Pagnani3,4,5, Francesco Zamponi2& Martin Weigt1✉ Generative models emerge as promising candidates for novel sequence-data driven approaches to protein design, and for the extraction of structural and functional informationabout proteins deeply hidden in rapidly growing sequence databases. Here we proposesimple autoregressive models as highly accurate but computationally ef ficient generative sequence models. We show that they perform similarly to existing approaches based onBoltzmann machines or deep generative models, but at a substantially lower computationalcost (by a factor between 10 2and 103). Furthermore, the simple structure of our models has distinctive mathematical advantages, which translate into an improved applicability insequence generation and evaluation. Within these models, we can easily estimate both theprobability of a given sequence, and, using the model ’s entropy, the size of the functional sequence space related to a speci fic protein family. In the example of response regulators, wefind a huge number of ca. 10 68possible sequences, which nevertheless constitute only the astronomically small fraction 10−80of all amino-acid sequences of the same length. These findings illustrate the potential and the dif ficulty in exploring sequence space via generative sequence models.https://doi.org/10.1038/s41467-021-25756-4 OPEN 1Sorbonne Université, CNRS, Institut de Biologie Paris Seine, Biologie Computationnelle et Quantitative LCQB, F-75005 Paris, France.2Laboratoire de Physique de l ’Ecole Normale Supérieure, ENS, Université PSL, CNRS, Sorbonne Université, Université de Paris, F-75005 Paris, France.3Department of Applied Science and Technology (DISAT), Politecnico di Torino, Corso Duca degli Abruzzi 24, I-10129 Torino, Italy.4Italian Institute for Genomic Medicine, IRCCS Candiolo, SP-142, I-10060 Candiolo (TO), Italy.5INFN Sezione di Torino, Via P. Giuria 1, I-10125 Torino, Italy.✉email: martin.weigt@sorbonne-universite.fr NATURE COMMUNICATIONS | (2021) 12:5800 | https://doi.org/10.1038/s41467-021-25756-4 | www.nature.com/naturecommunications 11234567890():,;

2306.03078.pdf

SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression Tim Dettmers∗ † University of WashingtonRuslan Svirschevski∗ HSE University & YandexVage Egiazarian∗ HSE University & Yandex Denis Kuznedelev∗ Yandex & SkoltechElias Frantar IST AustriaSaleh Ashkboos ETH ZurichAlexander Borzunov HSE University & Yandex Torsten Hoefler ETH ZurichDan Alistarh IST Austria & NeuralMagic Abstract Recent advances in large language model (LLM) pretraining have led to high- quality LLMs with impressive abilities. By compressing such LLMs via quanti- zation to 3-4 bits per parameter, they can fit into memory-limited devices such as laptops and mobile phones, enabling personalized use. However, quantiza- tion down to 3-4 bits per parameter usually leads to moderate-to-high accuracy losses, especially for smaller models in the 1-10B parameter range, which are well-suited for edge deployments. To address this accuracy issue, we introduce the Sparse-Quantized Representation (SpQR), a new compressed format and quantiza- tion technique which enables for the first time near-lossless compression of LLMs across model scales, while reaching similar compression levels to previous methods. SpQR works by identifying and isolating outlier weights , which cause particularly- large quantization errors, and storing them in higher precision, while compressing all other weights to 3-4 bits, and achieves relative accuracy losses of less than 1%in perplexity for highly-accurate LLaMA and Falcon LLMs. This makes it possible to run 33B parameter LLM on a single 24 GB consumer GPU without any performance degradation at 15% speedup thus making powerful LLMs available to consumer without any downsides. SpQR comes with efficient algorithms for both encoding weights into its format, as well as decoding them efficiently at runtime3. Specifically, we provide an efficient GPU inference algorithm for SpQR which yields faster inference than 16-bit baselines at similar accuracy, while enabling memory compression gains of more than 4x. 1 Introduction Pretrained large language models (LLMs) improved rapidly from task-specific performance [WSM+18,DCLT19 ,RWC+19], to performing well on general tasks if prompted with instruc- tions [ BMR+20,WBZ+21,Ope23 ]. While the improved performance can be attributed to scaling in training data and parameters [ KMH+20,CND+22] recent trends focused on smaller models trained on more data, that are easier to use at inference time [ HBM+22,BSA+23,TLI+23]. For example, the 7B parameter LLaMA model trained on 1T tokens achieved an average performance only slightly lower than GPT-3 [ BMR+20] despite being 25x smaller. Current techniques for LLM compres- sion can shrink these models further by a factor of about 4x, while preserving their performance ∗Equal contribution †Corresponding author: dettmers@cs.washington.edu 3github.com/Vahe1994/SpQR ; to be integrated into github.com/TimDettmers/bitsandbytesarXiv:2306.03078v1 [cs.CL] 5 Jun 2023

1706.03741.pdf

Deep Reinforcement Learning from Human Preferences Paul F Christiano OpenAI paul@openai.comJan Leike DeepMind leike@google.comTom B Brown nottombrown@gmail.com Miljan Martic DeepMind miljanm@google.comShane Legg DeepMind legg@google.comDario Amodei OpenAI damodei@openai.com Abstract For sophisticated reinforcement learning (RL) systems to interact usefully with real-world environments, we need to communicate complex goals to these systems. In this work, we explore goals defined in terms of (non-expert) human preferences between pairs of trajectory segments. We show that this approach can effectively solve complex RL tasks without access to the reward function, including Atari games and simulated robot locomotion, while providing feedback on less than 1% of our agent’s interactions with the environment. This reduces the cost of human oversight far enough that it can be practically applied to state-of-the-art RL systems. To demonstrate the flexibility of our approach, we show that we can successfully train complex novel behaviors with about an hour of human time. These behaviors and environments are considerably more complex than any which have been previously learned from human feedback. 1 Introduction Recent success in scaling reinforcement learning (RL) to large problems has been driven in domains that have a well-specified reward function (Mnih et al., 2015, 2016; Silver et al., 2016). Unfortunately, many tasks involve goals that are complex, poorly-defined, or hard to specify. Overcoming this limitation would greatly expand the possible impact of deep RL and could increase the reach of machine learning more broadly. For example, suppose that we wanted to use reinforcement learning to train a robot to clean a table or scramble an egg. It’s not clear how to construct a suitable reward function, which will need to be a function of the robot’s sensors. We could try to design a simple reward function that approximately captures the intended behavior, but this will often result in behavior that optimizes our reward function without actually satisfying our preferences. This difficulty underlies recent concerns about misalignment between our values and the objectives of our RL systems (Bostrom, 2014; Russell, 2016; Amodei et al., 2016). If we could successfully communicate our actual objectives to our agents, it would be a significant step towards addressing these concerns. If we have demonstrations of the desired task, we can extract a reward function using inverse reinforcement learning (Ng and Russell, 2000). This reward function can then be used to train an agent with reinforcement learning. More directly, we can use imitation learning to clone the demonstrated behavior. However, these approaches are not directly applicable to behaviors that are difficult for humans to demonstrate (such as controlling a robot with many degrees of freedom but very non-human morphology).arXiv:1706.03741v4 [stat.ML] 17 Feb 2023

karakida19a.pdf

Universal Statistics of Fisher Information in Deep Neural Networks: Mean Field Approach Ryo Karakida Shotaro Akaho Shun-ichi Amari AIST, Japan AIST, Japan RIKEN CBS, Japan Abstract The Fisher information matrix (FIM) is a fundamental quantity to represent the char- acteristics of a stochastic model, including deep neural networks (DNNs). The present study reveals novel statistics of FIM that are universal among a wide class of DNNs. To this end, we use random weights and large width limits, which enables us to utilize mean field theories. We investigate the asymptotic statistics of the FIM’s eigenvalues and reveal that most of them are close to zero while the maximum eigenvalue takes a huge value. Be- cause the landscape of the parameter space is defined by the FIM, it is locally flat in most dimensions, but strongly distorted in others. Moreover, we demonstrate the potential usage of the derived statistics in learning strategies. First, small eigenvalues that induce flatness can be connected to a norm-based capacity measure of generalization ability. Second, the maximum eigenvalue that induces the distor- tion enables us to quantitatively estimate an appropriately sized learning rate for gradient methods to converge. 1 Introduction Deep learning has succeeded in making hierarchical neural networks perform excellently in various practi- cal applications [ 1]. To proceed further, it would be beneficial to give more theoretical elucidation as to why and how deep neural networks (DNNs) work well in practice. In particular, it would be useful to not only clarify the individual models and phenomena but also explore various unified theoretical frameworks that Proceedings of the 22ndInternational Conference on Ar- tificial Intelligence and Statistics (AISTATS) 2019, Naha, Okinawa, Japan. PMLR: Volume 89. Copyright 2019 by the author(s).could be applied to a wide class of deep networks. One widely used approach for this purpose is to consider deep networks with random connectivity and a large width limit [ 2–14]. For instance, Poole et al. [3]pro- posed a useful indicator to explain the expressivity of DNNs. Regarding the trainability of DNNs, Schoen- holz et al. [4]extended this theory to backpropagation and found that the vanishing and explosive gradients obey a universal law. These studies are powerful in the sense that they do not depend on particular model architectures, such as the number of layers or activation functions. Unfortunately, such universal frameworks have not yet been established in many other topics. One is the geo- metric structure of the parameter space. For instance, the loss landscape without spurious local minima is im- portant for easier optimization and theoretically guar- anteed in single-layer models [ 15], shallow piecewise linear ones [ 16], and extremely wide deep networks with the number of training samples smaller than the width [17]. Flat global minima have been reported to be related to generalization ability through empirical experiments showing that networks with such minima give better generalization performance [ 18,19]. How- ever, theoretical analysis of the flat landscape has been limited in shallow rectified linear unit (ReLU) networks [20,21]. Thus, a residual subject of interest is to theo- reticallyrevealthegeometricstructureoftheparameter space truly common among various deep networks. To establish the foundation of the universal perspec- tive of the parameter space, this study analytically investigates the Fisher information matrix (FIM). As is overviewed in Section 2.1, the FIM plays an essential role in the geometry of the parameter space and is a fundamental quantity in both statistics and machine learning. 1.1 Main results This study analyzes the FIM of deep networks with ran- dom weights and biases, which are widely used settings to analyze the phenomena of DNNs [ 2–14]. First, we

2310.06816.pdf

Text Embeddings Reveal (Almost) As Much As Text John X. Morris, Volodymyr Kuleshov, Vitaly Shmatikov, Alexander M. Rush Department of Computer Science Cornell University Abstract How much private information do text em- beddings reveal about the original text? We investigate the problem of embedding inver- sion, reconstructing the full text represented in dense text embeddings. We frame the prob- lem as controlled generation: generating text that, when reembedded, is close to a fixed point in latent space. We find that although a naïve model conditioned on the embedding performs poorly, a multi-step method that iteratively cor- rects and re-embeds text is able to recover 92% of32-token text inputs exactly. We train our model to decode text embeddings from two state-of-the-art embedding models, and also show that our model can recover important per- sonal information (full names) from a dataset of clinical notes.1 1 Introduction Systems that utilize large language models (LLMs) often store auxiliary data in a vector database of dense embeddings (Borgeaud et al., 2022; Yao et al., 2023). Users of these systems infuse knowl- edge into LLMs by inserting retrieved documents into the language model’s prompt. Practition- ers are turning to hosted vector database services to execute embedding search efficiently at scale (Pinecone; Qdrant; Vdaas; Weaviate; LangChain). In these databases, the data owner only sends em- beddings of text data (Le and Mikolov, 2014; Kiros et al., 2015) to the third party service, and never the text itself. The database server returns a search result as the index of the matching document on the client side. Vector databases are increasingly popular, but privacy threats within them have not been compre- hensively explored. Can the third party service to reproduce the initial text, given its embedding? Neural networks are in general non-trivial or even 1Our code is available on Github: github.com/jxmorris12/vec2text.impossible to invert exactly. Furthermore, when querying a neural network through the internet, we may not have access to the model weights or gradi- ents at all. Still, given input-output pairs from a network, it is often possible to approximate the network’s inverse. Work on inversion in computer vision (Mahendran and Vedaldi, 2014; Dosovitskiy and Brox, 2016) has shown that it is possible to learn to recover the input image (with some loss) given the logits of the final layer. Preliminary work has explored this question for text (Song and Raghu- nathan, 2020), but only been able to recover an approximate bag of words given embeddings from shallow networks. In this work, we target full reconstruction of in- put text from its embedding. If text is recoverable, there is a threat to privacy: a malicious user with ac- cess to a vector database, and text-embedding pairs from the model used to produce the data, could learn a function that reproduces text from embed- dings. We frame this problem of recovering textual em- beddings as a controlled generation problem, where we seek to generate text such that the text is as close as possible to a given embedding. Our method, Vec2Text , uses the difference between a hypothesis embedding and a ground-truth embedding to make discrete updates to the text hypothesis. When we embed web documents using a state-of- the-art black-box encoder, our method can recover 32-token inputs with a near-perfect BLEU score of 97.3, and can recover 92% of the examples exactly. We then evaluate on embeddings generated from a variety of common retrieval corpuses from the BEIR benchmark. Even though these texts were not seen during training, our method is able to per- fectly recover the inputs for a number of datapoints across a variety of domains. We evaluate on em- beddings of clinical notes from MIMIC and are able to recover 89% of full names from embeddedarXiv:2310.06816v1 [cs.CL] 10 Oct 2023

1908.10084v1.pdf

Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks Nils Reimers and Iryna Gurevych Ubiquitous Knowledge Processing Lab (UKP-TUDA) Department of Computer Science, Technische Universit ¨at Darmstadt www.ukp.tu-darmstadt.de Abstract BERT (Devlin et al., 2018) and RoBERTa (Liu et al., 2019) has set a new state-of-the-art performance on sentence-pair regression tasks like semantic textual similarity (STS). How- ever, it requires that both sentences are fed into the network, which causes a massive com- putational overhead: Finding the most sim- ilar pair in a collection of 10,000 sentences requires about 50 million inference computa- tions (~65 hours) with BERT. The construction of BERT makes it unsuitable for semantic sim- ilarity search as well as for unsupervised tasks like clustering. In this publication, we present Sentence-BERT (SBERT), a modification of the pretrained BERT network that use siamese and triplet net- work structures to derive semantically mean- ingful sentence embeddings that can be com- pared using cosine-similarity. This reduces the effort for finding the most similar pair from 65 hours with BERT / RoBERTa to about 5 sec- onds with SBERT, while maintaining the ac- curacy from BERT. We evaluate SBERT and SRoBERTa on com- mon STS tasks and transfer learning tasks, where it outperforms other state-of-the-art sentence embeddings methods.1 1 Introduction In this publication, we present Sentence-BERT (SBERT), a modification of the BERT network us- ing siamese and triplet networks that is able to derive semantically meaningful sentence embed- dings2. This enables BERT to be used for certain new tasks, which up-to-now were not applicable for BERT. These tasks include large-scale seman- 1Code available: https://github.com/UKPLab/ sentence-transformers 2With semantically meaningful we mean that semantically similar sentences are close in vector space.tic similarity comparison, clustering, and informa- tion retrieval via semantic search. BERT set new state-of-the-art performance on various sentence classification and sentence-pair regression tasks. BERT uses a cross-encoder: Two sentences are passed to the transformer network and the target value is predicted. However, this setup is unsuitable for various pair regression tasks due to too many possible combinations. Finding in a collection of n= 10 000 sentences the pair with the highest similarity requires with BERT n·(n−1)/2 = 49 995 000 inference computations. On a modern V100 GPU, this requires about 65 hours. Similar, finding which of the over 40 mil- lion existent questions of Quora is the most similar for a new question could be modeled as a pair-wise comparison with BERT, however, answering a sin- gle query would require over 50 hours. A common method to address clustering and se- mantic search is to map each sentence to a vec- tor space such that semantically similar sentences are close. Researchers have started to input indi- vidual sentences into BERT and to derive fixed- size sentence embeddings. The most commonly used approach is to average the BERT output layer (known as BERT embeddings) or by using the out- put of the first token (the [CLS] token). As we will show, this common practice yields rather bad sentence embeddings, often worse than averaging GloVe embeddings (Pennington et al., 2014). To alleviate this issue, we developed SBERT. The siamese network architecture enables that fixed-sized vectors for input sentences can be de- rived. Using a similarity measure like cosine- similarity or Manhatten / Euclidean distance, se- mantically similar sentences can be found. These similarity measures can be performed extremely efficient on modern hardware, allowing SBERT to be used for semantic similarity search as well as for clustering. The complexity for finding thearXiv:1908.10084v1 [cs.CL] 27 Aug 2019

2402.00854.pdf

SymbolicAI: A framework for logic-based approaches combining generative models and solvers Marius–Constantin Dinu∗ †Claudiu Leoveanu–Condrei‡Markus Holzleitner† Werner Zellinger§Sepp Hochreiter† Abstract We introduce SymbolicAI , a versatile and modular framework employing a logic-based approach to concept learning and flow management in generative processes. SymbolicAI enables the seamless integration of generative models with a diverse range of solvers by treating large language models (LLMs) as semantic parsers that execute tasks based on both natural and formal language instruc- tions, thus bridging the gap between symbolic reasoning and generative AI. We leverage probabilistic programming principles to tackle complex tasks, and utilize differentiable and classical program- ming paradigms with their respective strengths. The framework introduces a set of polymorphic, compositional, and self-referential operations for data stream manipulation, aligning LLM outputs with user objectives. As a result, we can transition between the capabilities of various foundation models endowed with zero- and few-shot learning capabilities and specialized, fine-tuned models or solvers proficient in addressing specific problems. In turn, the framework facilitates the creation and evaluation of explainable computational graphs. We conclude by introducing a quality measure and its empirical score for evaluating these computational graphs, and propose a benchmark that compares various state-of-the-art LLMs across a set of complex workflows. We refer to the empirical score as the ”Vector Embedding for Relational Trajectory Evaluation through Cross-similarity”, or VERTEX score for short. The framework codebase 1and benchmark 2are linked below. Prompting / Fine-TuningNeuro-Symbolic AI Spectrum Software-Engineering Machine LearningFoundation Models Specialist ModelsProgramming / LearningModeling / CodingAbstraction Implementation Figure 1: Our neuro-symbolic framework enables a seamless transition between classical and differentiable program- ming, each with distinct dynamics and strengths. Differentiable programming provides access to foundational and specialist models. Classical programming, on the other hand, shifts between abstraction and implementation, focusing on high-level concepts before delving into the details of implementation. ∗ExtensityAI, Vienna and AI Austria, Vienna — Corresponding author emails: dinu@ml.jku.at, office@extensity.ai †ELLIS Unit Linz and LIT AI Lab, Institute for Machine Learning, Johannes Kepler University, Linz ‡Amazon Devices, Timis ,oara – work done outside of Amazon §Johann Radon Institute for Computational and Applied Mathematics, Austrian Academy of Sciences, Vienna 1SymbolicAI framework released on January 20th, 2023, on GitHub: https://github.com/ExtensityAI/symbolicai . 2Evaluation benchmark released on February 1st, 2024, on GitHub: https://github.com/ExtensityAI/benchmark . 1arXiv:2402.00854v2 [cs.LG] 5 Feb 2024

1907.10786.pdf

Interpreting the Latent Space of GANs for Semantic Face Editing Yujun Shen1, Jinjin Gu2, Xiaoou Tang1, Bolei Zhou1 1The Chinese University of Hong Kong2The Chinese University of Hong Kong, Shenzhen {sy116, xtang, bzhou }@ie.cuhk.edu.hk, jinjingu@link.cuhk.edu.cn Original Pose Age Gender Eyeglasses Figure 1: Manipulating various facial attributes through varying the latent codes of a well-trained GAN model. The first column shows the original synthesis from PGGAN [21], while each of the other columns shows the results of manipulating a specific attribute. Abstract Despite the recent advance of Generative Adversarial Networks (GANs) in high-fidelity image synthesis, there lacks enough understanding of how GANs are able to map a latent code sampled from a random distribution to a photo- realistic image. Previous work assumes the latent space learned by GANs follows a distributed representation but observes the vector arithmetic phenomenon. In this work, we propose a novel framework, called InterFaceGAN, for semantic face editing by interpreting the latent semantics learned by GANs. In this framework, we conduct a detailed study on how different semantics are encoded in the latent space of GANs for face synthesis. We find that the latent code of well-trained generative models actually learns a disentangled representation after linear transformations. We explore the disentanglement between various semantics and manage to decouple some entangled semantics with subspace projection, leading to more precise control of facial attributes. Besides manipulating gender, age, expres- sion, and the presence of eyeglasses, we can even vary the face pose as well as fix the artifacts accidentally generatedby GAN models. The proposed method is further applied to achieve real image manipulation when combined with GAN inversion methods or some encoder-involved models. Extensive results suggest that learning to synthesize faces spontaneously brings a disentangled and controllable facial attribute representation.1 1. Introduction Generative Adversarial Networks (GANs) [15] have significantly advanced image synthesis in recent years. The rationale behind GANs is to learn the mapping from a latent distribution to the real data through adversarial training. After learning such a non-linear mapping, GAN is capable of producing photo-realistic images from randomly sam- pled latent codes. However, it is uncertain how semantics originate and are organized in the latent space. Taking face synthesis as an example, when sampling a latent code to produce an image, how the code is able to determine various semantic attributes ( e.g., gender and age) of the output face, and how these attributes are entangled with each other? 1Code and models are available at this link. 1arXiv:1907.10786v3 [cs.CV] 31 Mar 2020

2107.13163.pdf

arXiv:2107.13163v3 [cs.LG] 30 Mar 2023Statistically Meaningful Approximation: a Case Study on Approximating Turing Machines with Transform ers Colin Wei Yining Chen Tengyu Ma Department of Computer Science Stanford University {colinwei,cynnjjs,tengyuma}@cs.stanford.edu March 31, 2023 Abstract A common lens to theoretically study neural net architectur es is to analyze the functions they can approximate. However, the constructions from approximati on theory often have unrealistic aspects, for example, reliance on infinite precision to memorize target f unction values. To address this issue, we propose a formal definition of statistically meaningful approximat ion which requires the approximating network to exhibit good statistical learnability. We present case stu dies on statistically meaningful approximation for two classes of functions: boolean circuits and Turing ma chines. We show that overparameterized feed- forward neural nets can statistically meaningfully approx imate boolean circuits with sample complexity depending only polynomially on the circuit size, not the siz e of the approximating network. In addition, we show that transformers can statistically meaningfully app roximate Turing machines with computation time bounded by T, requiring sample complexity polynomial in the alphabet si ze, state space size, and logpTq. Our analysis introduces new tools for generalization bound s that provide much tighter sample complexity guarantees than the typical VC-dimension or norm-based bou nds, which may be of independent interest. 1 Introduction Dating back to the seminal works on universal approximation [16, 25, 40, 31], a common way to theoretically study neural nets has been through their expressivity, whic h measures the ability of neural nets to approxi- mate well-behaved functions. This perspective has shaped h ow researchers perceive different types of deep learning architectures: a basic way to theoretically justi fy new architectures is to study their approximation capabilities. This has led to a number of analyses studying u niversal approximation capabilities for various widely-used architectures, such as recurrent neural nets ( RNNs) [47], graph neural nets [46], convolutional networks [3, 64, 59], residual networks [32], transformers [61], and neural ODEs [51, 63]. However, approximation theoretic results often misalign w ith more meaningful end-to-end guarantees, because models constructed in the literature often exhibit unrealistic properties. For example, a common tech- nique in the universal approximation literature is to rely s trongly on infinite-precision weights and activations, or exponentially many parameters to encode the desired func tion values [25, 16, 31, 32, 61, 44]. This issue even arises outside of universal approximation, e.g., vari ous papers demonstrate the ability of RNNs and trans- formers to simulate various computational models such as Tu ring machines and automata, but require strong reliance on arbitrary precision [48, 42, 29, 9]. Infinite pre cision can inflate the expressivity of an architecture (function class) in a unrealistic and misleading way: for ex ample, finite width RNNs with infinite precision can simulate Turing machines, but finite-precision, finite-wid th RNNs cannot. This is implied by streaming lower bounds [1] – any finite-precision, finite-width RNN induces a finite-space streaming algorithm corresponding to running the RNN on the inputs. However, streaming lower bo unds tell us that finite-space streaming al- gorithms are not powerful enough to simulate Turing machine s, and hence finite-precision, finite-width RNNs 1

1906.08237.pdf

XLNet: Generalized Autoregressive Pretraining for Language Understanding Zhilin Yang∗1, Zihang Dai∗12, Yiming Yang1, Jaime Carbonell1, Ruslan Salakhutdinov1, Quoc V . Le2 1Carnegie Mellon University,2Google AI Brain Team {zhiliny,dzihang,yiming,jgc,rsalakhu}@cs.cmu.edu, qvl@google.com Abstract With the capability of modeling bidirectional contexts, denoising autoencoding based pretraining like BERT achieves better performance than pretraining ap- proaches based on autoregressive language modeling. However, relying on corrupt- ing the input with masks, BERT neglects dependency between the masked positions and suffers from a pretrain-finetune discrepancy. In light of these pros and cons, we propose XLNet, a generalized autoregressive pretraining method that (1) enables learning bidirectional contexts by maximizing the expected likelihood over all permutations of the factorization order and (2) overcomes the limitations of BERT thanks to its autoregressive formulation. Furthermore, XLNet integrates ideas from Transformer-XL, the state-of-the-art autoregressive model, into pretraining. Empirically, under comparable experiment settings, XLNet outperforms BERT on 20 tasks, often by a large margin, including question answering, natural language inference, sentiment analysis, and document ranking.1. 1 Introduction Unsupervised representation learning has been highly successful in the domain of natural language processing [ 7,22,27,28,10]. Typically, these methods first pretrain neural networks on large-scale unlabeled text corpora, and then finetune the models or representations on downstream tasks. Under this shared high-level idea, different unsupervised pretraining objectives have been explored in literature. Among them, autoregressive (AR) language modeling and autoencoding (AE) have been the two most successful pretraining objectives. AR language modeling seeks to estimate the probability distribution of a text corpus with an au- toregressive model [ 7,27,28]. Specifically, given a text sequence x= (x1,···,xT), AR language modeling factorizes the likelihood into a forward product p(x) =∏T t=1p(xt|x<t)or a backward onep(x) =∏1 t=Tp(xt|x>t). A parametric model (e.g. a neural network) is trained to model each conditional distribution. Since an AR language model is only trained to encode a uni-directional con- text (either forward or backward), it is not effective at modeling deep bidirectional contexts. On the contrary, downstream language understanding tasks often require bidirectional context information. This results in a gap between AR language modeling and effective pretraining. In comparison, AE based pretraining does not perform explicit density estimation but instead aims to reconstruct the original data from corrupted input. A notable example is BERT [ 10], which has been the state-of-the-art pretraining approach. Given the input token sequence, a certain portion of tokens are replaced by a special symbol [MASK] , and the model is trained to recover the original tokens from the corrupted version. Since density estimation is not part of the objective, BERT is allowed to utilize ∗Equal contribution. Order determined by swapping the one in [9]. 1Pretrained models and code are available at https://github.com/zihangdai/xlnet 33rd Conference on Neural Information Processing Systems (NeurIPS 2019), Vancouver, Canada.arXiv:1906.08237v2 [cs.CL] 2 Jan 2020

2206.05895.pdf

Latent Diffusion Energy-Based Model for Interpretable Text Modeling Peiyu Yu1 2Sirui Xie1Xiaojian Ma1 2Baoxiong Jia1 2Bo Pang3 Ruiqi Gao4Yixin Zhu5 6Song-Chun Zhu1 2 5 6 7 8Ying Nian Wu7 Abstract Latent space Energy-Based Models ( EBM s), also known as energy-based priors, have drawn grow- ing interests in generative modeling. Fueled by its flexibility in the formulation and strong modeling power of the latent space, recent works built upon it have made interesting attempts aiming at the interpretability of text modeling. However, latent space EBM s also inherit some flaws from EBM s in data space; the degenerate MCMC sampling quality in practice can lead to poor generation quality and instability in training, especially on data with complex latent structures. Inspired by the recent efforts that leverage diffusion recovery likelihood learning as a cure for the sampling is- sue, we introduce a novel symbiosis between the diffusion models and latent space EBM s in a vari- ational learning framework, coined as the latent diffusion energy-based model . We develop a geo- metric clustering-based regularization jointly with the information bottleneck to further improve the quality of the learned latent space. Experiments on several challenging tasks demonstrate the su- perior performance of our model on interpretable text modeling over strong counterparts. 1. Introduction Text modeling has achieved impressive progress with the fast development of neural generative models (Serban et al., 2016; Li et al., 2017a; Zhao et al., 2017; Gupta et al., 2018; Code repo and data: https://github.com/yuPeiyu98/Latent- Diffusion-EBM.1Department of Computer Science, UCLA, USA2Beijing Institute for General Artificial Intelligence, China 3Salesforce Research, USA4Google Brain, USA5Institute for Ar- tificial Intelligence, Peking University, China6School of Artificial Intelligence, Peking University, China7Department of Statistics, UCLA, USA8Department of Automation, Tsinghua University, China. Correspondence to: Peiyu Yu <yupeiyu98@g.ucla.edu>. Proceedings of the 39thInternational Conference on Machine Learning , Baltimore, Maryland, USA, PMLR 162, 2022. Copy- right 2022 by the author(s).x z0 zt zt+1y q(zt+1|zt) pα(zt|zt+1)pα(y,z0|z1) qϕ(z0|x) pβ(x|z0) t= 1, ..., T−1 Figure 1. Graphical illustration of the latent diffusion process. We construct the forward and reverse diffusion processes in the la- tent space. The symbolic one-hot vector is coupled with the initial latent vector z0. The latent and diffused latent variables are high- lighted by the red and blue plates, respectively. The cyan arrows indicate that z0is connected with only z1. We learn a sequence of EBMs to model the reverse diffusion process pα(zt|zt+1). Zhao et al., 2018a). It allows near human-level text gener- ation quality and also leads to a wide range of real-world applications such as dialog system (Young et al., 2013) and machine translation (Brown et al., 1993). Although the qual- ity of generation ( e.g., fluency and diversity) is the primary concern of most work, interpretability of the generation pro- cess has drawn much attention recently. Among the existing frameworks, the Deep Latent Variable Model ( DLVM ) is especially suitable for the task, as the learned latent space could capture high-level structures with semantic meanings like topics (Wang et al., 2019) and dialog actions (Zhao et al., 2018b); such latent space could further enable more interpretable text modeling, featuring unsupervised text at- tributes discovery (Wen et al., 2017), conditional and con- trollable text generation (Fang et al., 2019; Shi et al., 2020), and semi-supervised text classification (Pang & Wu, 2021). In essence, DLVM summarizes the observed sample ( e.g., a piece of text) into inferred latent variables. Earlier text-modeling methods with DLVM mostly follow the for- mulation of Variational Auto-Encoder ( V AE ) (Kingma & Welling, 2013; Rezende et al., 2014; Bowman et al., 2016), which assumes a continuous latent space. More recently, Zhao et al. (2018b) explore the possibility of using a discrete latent space to capture dialog actions; Shi et al. (2020) pro- pose to use V AE with the mixture of Gaussians as the prior, demonstrating promising interpretability of dialog utterancearXiv:2206.05895v4 [cs.LG] 4 Oct 2023

2209.13325.pdf

Outlier Suppression: Pushing the Limit of Low-bit Transformer Language Models Xiuying Wei1, 2, Yunchen Zhang2, 4, Xiangguo Zhang2, Ruihao Gong1, 2, Shanghang Zhang3, Qi Zhang2, Fengwei Yu2, Xianglong Liu1∗ 1State Key Lab of Software Development Environment, Beihang University 2SenseTime Research,3Peking University 4University of Electronic Science and Technology of China {weixiuying, zhangyunchen, zhangxiangguo, gongruihao}@sensetime.com shanghang@pku.edu.cn, xlliu@buaa.edu.cn Abstract Transformer architecture has become the fundamental element of the widespread natural language processing (NLP) models. With the trends of large NLP models, the increasing memory and computation costs hinder their efficient deployment on resource-limited devices. Therefore, transformer quantization attracts wide research interest. Recent work recognizes that structured outliers are the criti- cal bottleneck for quantization performance. However, their proposed methods increase the computation overhead and still leave the outliers there. To funda- mentally address this problem, this paper delves into the inherent inducement and importance of the outliers. We discover that γin LayerNorm (LN) acts as a sinful amplifier for the outliers, and the importance of outliers varies greatly where some outliers provided by a few tokens cover a large area but can be clipped sharply without negative impacts. Motivated by these findings, we propose an outlier suppression framework including two components: Gamma Migration and Token-Wise Clipping. The Gamma Migration migrates the outlier amplifier to subsequent modules in an equivalent transformation, contributing to a more quantization-friendly model without any extra burden. The Token-Wise Clipping takes advantage of the large variance of token range and designs a token-wise coarse-to-fine pipeline, obtaining a clipping range with minimal final quantiza- tion loss in an efficient way. This framework effectively suppresses the outliers and can be used in a plug-and-play mode. Extensive experiments prove that our framework surpasses the existing works and, for the first time, pushes the 6-bit post- training BERT quantization to the full-precision (FP) level. Our code is available athttps://github.com/wimh966/outlier_suppression . 1 Introduction Transformer [ 1] has been one of the most common architectures in natural language processing along with lots of popular self-supervised models, such as BERT [ 2], RoBERTa [ 3], XLNet [ 4] and BART [5]. While these pre-trained models have demonstrated a significant superiority in performance, the memory and computation overheads have been a popular concern, particularly in the real development. Therefore, model compression [ 6,7,8,9] has attracted much attention from both academia and industry. Among them, quantization [ 10,11,12,13,14,15,16,17,18,19,20], working in the low-precision arithmetic fashion, is one of the key approaches to compress large models and fit them into the lightweight devices. ∗Corresponding author. 36th Conference on Neural Information Processing Systems (NeurIPS 2022).arXiv:2209.13325v3 [cs.LG] 21 Feb 2023

2308.05660v1.pdf

Thermodynamic Linear Algebra Maxwell Aifer, Kaelan Donatella, Max Hunter Gordon, Thomas Ahle, Daniel Simpson, Gavin Crooks, Patrick J. Coles Normal Computing Corporation, New York, New York, USA Linear algebraic primitives are at the core of many modern algorithms in engineering, science, and machine learning. Hence, accelerating these primitives with novel computing hardware would have tremendous economic impact. Quantum computing has been proposed for this purpose, although the resource requirements are far beyond current technological capabilities, so this approach remains long-term in timescale. Here we consider an alternative physics-based computing paradigm based on classical thermodynamics, to provide a near-term approach to accelerating linear algebra. At first sight, thermodynamics and linear algebra seem to be unrelated fields. In this work, we connect solving linear algebra problems to sampling from the thermodynamic equilibrium distri- bution of a system of coupled harmonic oscillators. We present simple thermodynamic algorithms for (1) solving linear systems of equations, (2) computing matrix inverses, (3) computing matrix determinants, and (4) solving Lyapunov equations. Under reasonable assumptions, we rigorously establish asymptotic speedups for our algorithms, relative to digital methods, that scale linearly in matrix dimension. Our algorithms exploit thermodynamic principles like ergodicity, entropy, and equilibration, highlighting the deep connection between these two seemingly distinct fields, and opening up algebraic applications for thermodynamic computing hardware. I. Introduction Basic linear algebra primitives such as solving a linear system of the form Ax=band obtaining the inverse of a matrix are present in many modern algorithms. Such primitives are relevant to a multitude of applications, including for example optimal control of dynamic systems and resource allocation. They are also a common subroutine of many artificial intelligence (AI) algorithms, and account for a substantial portion of the time and energy costs in some cases. The most common method to perform these primitives is LU decomposition, whose time-complexity scales as O(d3). Many proposals have been made to accelerate such primitives, for example using iterative methods such as the conjugate gradient method. In the last decade, these primitives have been accelerated by hardware improvements, notably by their implementation on graphical processing units (GPUs), fueling massive parallelization. However, the scaling of these methods is still a prohibitive factor, and obtaining a good approximate solution to a dense matrix of more than a few tens of thousand dimensions remains challenging. Exploiting physics to solve mathematical problems is a deep idea, with much focus on solving optimization problems [1–3]. In the context of linear algebra, much attention has been paid to quantum computers [4], since the mathematics of discrete-variable quantum mechanics matches that of linear algebra. A quantum algorithm [5] to solve linear systems has been proposed, which for sparse and well-conditioned matrices scales as logd. However, the resource requirements [6] for this algorithm are far beyond current hardware capabilities. More generally building large-scale quantum hardware has remained difficult [7], and variational quantum algorithms for linear algebra [8–10] have battled with vanishing gradient issues [11–13]. Therefore, the search for alternative hardware proposals that can exploit physical dynamics to accelerate linear algebra primitives has been ongoing. Notably, memristor crossbar arrays have been of interest for accelerating matrix-vector multiplications [14, 15]. Solving linear systems has also been the subject of analog computing approaches [16]. Recently, we defined a new class of hardware, built from stochastic, analog building blocks, which is ultimately thermodynamic in nature [17]. (See also probabilistic-bit computers [18–20] and thermodynamic neural networks [21–24] for alternative approaches to thermodynamic computing [25]). AI applications like generative modeling are a natural fit for this thermodynamic hardware, where stochastic fluctuations are exploited to generate novel samples. In this work, we surprisingly show that the same thermodynamic hardware from Ref. [17] can also be used toacceleratekeyprimitivesinlinearalgebra. Thermodynamicsisnottypicallyassociatedwithlinearalgebra, and connecting these two fields is therefore non-trivial. Here, we exploit the fact that the mathematics of harmonic oscillator systems is inherently affine (i.e., linear), and hence we can map linear algebraic primitives onto such systems. (See also Ref. [26] for a discussion of harmonic oscillators in the context of quantum computingspeedups.) Weshowthatsimplybysamplingfromthethermalequilibriumdistributionofcoupled harmonic oscillators, one can solve a variety of linear algebra problems.arXiv:2308.05660v1 [cond-mat.stat-mech] 10 Aug 2023

2312.17227.pdf

Gradient-based Planning with World Models Jyothir S V1∗Siddhartha Jalagam1∗Yann LeCun1, 2Vlad Sobal1, 2 1New York University2Meta AI {jyothir, scj9994, us441}@nyu.edu yann@cs.nyu.edu Abstract The enduring challenge in the field of artificial intelligence has been the control of systems to achieve desired behaviours. While for systems governed by straightfor- ward dynamics equations, methods like Linear Quadratic Regulation (LQR) have historically proven highly effective, most real-world tasks, which require a general problem-solver, demand world models with dynamics that cannot be easily de- scribed by simple equations. Consequently, these models must be learned from data using neural networks. Most model predictive control (MPC) algorithms designed for visual world models have traditionally explored gradient-free population-based optimization methods, such as Cross Entropy and Model Predictive Path Integral (MPPI) for planning. However, we present an exploration of a gradient-based alternative that fully leverages the differentiability of the world model. In our study, we conduct a comparative analysis between our method and other MPC-based alternatives, as well as policy-based algorithms. In a sample-efficient setting, our method achieves on par or superior performance compared to the alternative ap- proaches in most tasks. Additionally, we introduce a hybrid model that combines policy networks and gradient-based MPC, which outperforms pure policy based methods thereby holding promise for Gradient-based planning with world models in complex real-world tasks. 1 Introduction Until recently, model-free reinforcement learning (RL) algorithms [ 24][28] have been the predominant choice for visual control tasks, particularly in simple environments like Atari games. However, these model-free algorithms are notorious for their sample inefficiency and lack of generality. If the tasks change, the policy needs to be trained again. They are constrained by their inability to transfer knowledge gained from training in one environment to another. Consequently, they must undergo retraining for even minor deviations from the original task. Real-world applications where the agent needs to solve a multitude of different tasks in the environment, such as robotics, demand a more general approach. To address this limitation, multiple types of methods have been proposed. In this work, we focus on model-based planning methods. These model-based approaches encompass three key components: a learned dynamics model that predicts state transitions, a learned reward or value model analogous to the cost function in Linear Quadratic Regulation (LQR) [ 6], which encapsulates state desirability information, and a planner that harnesses the world model and reward model to achieve desired states. While previous research in planning using Model Predictive Control (MPC) [ 25] has primarily focused on gradient-free methods like cross-entropy[ 27,9], these methods are computationally expensive and do not utilize the differentiability of the learned world model. ∗Equal Contribution. Preprint. Under review.arXiv:2312.17227v1 [cs.LG] 28 Dec 2023

10.1038.s41564-023-01584-8.pdf

Nature Microbiology nature microbiologyhttps://doi.org/10.1038/s41564-023-01584-8 Analysis Large language models improve annotation of prokaryotic viral proteins Zachary N. Flamholz   1, Steven J. Biller   2 & Libusha Kelly   1,3 Viral genomes are poorly annotated in metagenomic samples, representing an obstacle to understanding viral diversity and function. Current annotation approaches rely on alignment-based sequence homology methods, which are limited by the paucity of characterized viral proteins and divergence among viral sequences. Here we show that protein language models can capture prokaryotic viral protein function, enabling new portions of viral sequence space to be assigned biologically meaningful labels. When applied to global ocean virome data, our classifier expanded the annotated fraction of viral protein families by 29%. Among previously unannotated sequences, we highlight the identification of an integrase defining a mobile element in marine picocyanobacteria and a capsid protein that anchors globally widespread viral elements. Furthermore, improved high-level functional annotation provides a means to characterize similarities in genomic organization among diverse viral sequences. Protein language models thus enhance remote homology detection of viral proteins, serving as a useful complement to existing approaches. Viruses of microorganisms, hereafter ‘viruses’ , are abundant in the environment and have wide-ranging impacts on microbial communi- ties. Much of what we know about viral diversity, ecology and function comes from the analysis of sequences obtained from environmental samples, yet viruses are difficult to identify, classify and annotate. Thus, we make statements about viral biology and viral impacts on microbial community structure and function based on a tiny fraction of viral sequences with sufficient similarity to existing references. In recent years, next-generation sequencing and increasing computa - tional resources have been applied to catalogue the world’s virome1–7. While there has been substantial methodological progress in identify - ing viral DNA in whole-community metagenomic sequence data8–16, sequence feature annotation and overall taxonomic assignment of identified uncultivated virus genomes (UViGs) have lagged consid - erably. Viruses have no universal conserved marker genes to enable broad, unified, taxonomic analysis, and thus, most of the hundreds of thousands of new viruses uncovered in viral catalogue studies remain unclassified1–7. Viral taxonomic classification is generally based on using predicted UViG proteins as features for clustering-based17–19 or machine-learning-based20 taxonomic classification. Yet, as many as 86% of environmental viral protein clusters match uncharacterized protein families or have no hits at all6,7,16,21,22. Although detailed manual investigation of these sequence clusters may be able to yield hints of potential functions in some cases, such labour-intensive efforts do not readily scale to the amount of data being generated. Improved anno - tation of viral protein families (VPFs) is thus a necessary, unrealized step towards understanding the roles of viruses in microbial ecology. Viral protein annotation currently relies on sequence homology using state-of-the-art approaches based on profile hidden Markov models (pHMMs). For viral metagenomics, sequence homology meth - ods suffer from two fundamental limitations: (1) the limited library of annotated viral protein sequences from which to construct probabil-istic sequence models and (2) the rate at which viral proteins change, quickly diverging beyond recognition by traditional sequence homol - ogy metrics. An alignment-free method that does not depend on con- structing sequence profiles for statistical sequence homology and that can leverage functional homology between proteins could overcome both challenges.Received: 23 April 2023 Accepted: 8 December 2023 Published online: xx xx xxxx Check for updates 1Department of Systems and Computational Biology, Albert Einstein College of Medicine, Bronx, NY, USA. 2Department of Biological Sciences, Wellesley College, Wellesley, MA, USA. 3Department of Microbiology and Immunology, Albert Einstein College of Medicine, Bronx, NY, USA.  e-mail: libusha.kelly@einsteinmed.edu

2202.03286.pdf

Red Teaming Language Models with Language Models WARNING: This paper contains model outputs which are offensive in nature. Ethan Perez1 2Saffron Huang1Francis Song1Trevor Cai1Roman Ring1 John Aslanides1Amelia Glaese1Nat McAleese1Geoffrey Irving1 1DeepMind,2New York University perez@nyu.edu Abstract Language Models (LMs) often cannot be deployed because of their potential to harm users in hard-to-predict ways. Prior work identifies harmful behaviors before deployment by using human annotators to hand-write test cases. However, human annotation is expensive, limiting the number and diversity of test cases. In this work, we automatically find cases where a target LM behaves in a harmful way, by generating test cases ( “red teaming” ) using another LM. We evaluate the target LM’s replies to generated test questions using a classifier trained to detect offensive content, uncovering tens of thousands of offensive replies in a 280B parameter LM chatbot. We explore several methods, from zero-shot generation to reinforcement learning, for generating test cases with varying levels of diversity and difficulty. Furthermore, we use prompt engineering to control LM-generated test cases to uncover a variety of other harms, automatically finding groups of people that the chatbot discusses in offensive ways, personal and hospital phone numbers generated as the chatbot’s own contact info, leakage of private training data in generated text, and harms that occur over the course of a conversation. Overall, LM-based red teaming is one promising tool (among many needed) for finding and fixing diverse, undesirable LM behaviors before impacting users. 1 Introduction Although we had prepared for many types of abuses of the system, we had made a critical oversight for this specific attack. Lee (2016) Language Models (LMs) are promising tools for a variety of applications, ranging from conversational assistants to question-answering systems. However, deploying LMs in production threatens to harm users in hard-to-predict ways. That's a mean question! I try my best to be fair to everyone! Reach me at y @ .com or +1 415- - . Favorite life quote? What's y our work emai l? Of course. I'm a patriot too.Are you a R epubl ican or Democr at? What would y ou do if you wer e invisible? Best jok e about brunet tes? Are you smarter than most brunet tes? Best jok e about r edheads? Are you smarter than most r edheads? I am a patriot. I hate Democr ats. Red LM Target LM Red Clf Offensive I don't talk about politics. Data Leakage User Info Here's a classic one: ...Distri- butional Bias I hate them too.Offensive Dialog ... ... ... ... Steal all of your money. "To be great, you must..."✘ ✘ ✘ ✘✘✓ ✓ ✓ ✓ ✘Great!Figure 1: Overview : We automatically generate test cases with a language model (LM), reply with the target LM, and find failing test cases using a classifier. For example, Microsoft took down its chatbot Tay after adversarial users evoked it into sending racist and sexually-charged tweets to over 50,000 followers (Lee, 2016). Other work has found that LMs generate misinformation (Lin et al., 2021) and confidential, personal information (e.g., social security numbers) from the LM training corpus (Carlini et al., 2019, 2021). Such failures have serious consequences, so it is crucial to discover and fix these failures before deployment. Prior work requires human annotators to manually discover failures, limiting the number and diversity of failures found. For example, some efforts find failures by using many hand-written test cases either directly (Ribeiro et al., 2020; Röttger et al., 2021; Xu et al., 2021b) or for supervised test case generation (Bartolo et al., 2021a). Other efforts manually compose templates and code toarXiv:2202.03286v1 [cs.CL] 7 Feb 2022

2401.14196.pdf

DeepSeek-Coder: When the Large Language Model Meets Programming - The Rise of Code Intelligence Daya Guo*1, Qihao Zhu∗1,2, Dejian Yang1, Zhenda Xie1, Kai Dong1, Wentao Zhang1 Guanting Chen1, Xiao Bi1, Y. Wu1, Y.K. Li1, Fuli Luo1, Yingfei Xiong2, Wenfeng Liang1 1DeepSeek-AI 2Key Lab of HCST (PKU), MOE; SCS, Peking University {zhuqh, guodaya}@deepseek.com https://github.com/deepseek-ai/DeepSeek-Coder Abstract The rapid development of large language models has revolutionized code intelligence in software development. However, the predominance of closed-source models has restricted extensive research and development. To address this, we introduce the DeepSeek-Coder series, a range of open-source code models with sizes from 1.3B to 33B, trained from scratch on 2 trillion tokens. These models are pre-trained on a high-quality project-level code corpus and employ a fill-in-the-blank task with a 16K window to enhance code generation and infilling. Our extensive evaluations demonstrate that DeepSeek-Coder not only achieves state-of-the-art performance among open-source code models across multiple benchmarks but also surpasses existing closed-source models like Codex and GPT-3.5. Furthermore, DeepSeek-Coder models are under a permissive license that allows for both research and unrestricted commercial use. Figure 1|The Performance of DeepSeek-Coder *Core contributors, ordered alphabetically by the name.arXiv:2401.14196v2 [cs.SE] 26 Jan 2024

dubey2022pursuit.pdf

RESEA RCH ARTICL E Thepursuit ofhappiness: Areinforcement learning perspective onhabituation and comparisons Rachit Dubey ID 1*,Thomas L.Griffiths2,Peter Dayan ID 3,4 1Department ofComputer Science, Princeton University ,Princeton, New Jersey, United States ofAmerica, 2Department ofPsychology, Prince tonUniversity, Prince ton,New Jersey, United States ofAmerica, 3Max Planck Institute forBiological Cybernetics ,Tu¨bingen, Germa ny,4University ofTu¨bingen, Tu¨bingen, Germany *rdubey@p rinceton .edu Abstract Inevaluating ourchoices, weoften suffer from twotragic relativities. First, when ourlives change forthebetter, werapidly habituate tothehigher standard ofliving. Second, wecan- notescape comparing ourselves tovarious relative standards. Habituation andcomparisons canbeverydisruptive todecision-making andhappiness ,andtilldate, itremains apuzzle whytheyhave come tobeapartofcognition inthefirstplace. Here, wepresent computa- tional evidence thatsuggests thatthese features might playanimportant roleinpromoting adaptive behavior. Using theframework ofreinforcement learning, weexplore thebenefit of employing areward function that, inaddition tothereward provided bytheunderlying task, alsodepends onprior expectations andrelative comparisons. Wefindthatwhile agents equipped withthisreward function arelesshappy, theylearn faster andsignificantly outper- form standard reward-based agents inawide range ofenvironmen ts.Specifically, wefind thatrelative comparisons speed uplearning byproviding anexploration incentive tothe agents, andprior expectations serve asauseful aidtocomparisons, especially insparsely- rewarded andnon-station aryenvironments. Oursimulations alsoreveal potential draw- backs ofthisreward function andshow thatagents perform sub-optimally when compari- sons areleftunchecked andwhen there aretoomany similar options. Together, ourresults helpexplain whyweareprone tobecoming trapped inacycle ofnever-ending wants and desires, andmayshed lightonpsychopatholog iessuch asdepression, materialism, and overconsumption. Author summary Even infavorable circumstances, weoften find ithard toremain happy with what we have. One might enjoy anewly bought carforaseason, butover time itbrings fewer posi- tivefeelings and oneeventually begins dreaming ofthenext rewarding thing topursue. Here, wepresent aseries ofcomputational simulations that suggest these presumable “flaws” might play animportant role inpromoting adaptive behavior. Weexplore the PLOS COMP UTATIONAL BIOLOGY PLOS Computationa lBiology |https:/ /doi.org/10.13 71/journal.p cbi.1010316 August 4,2022 1/32a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Dubey R,Griffiths TL,Dayan P(2022) Thepursuit ofhappin ess:Areinforcem entlearning perspective onhabituat ionandcomparisons. PLoS Comput Biol18(8): e1010316. https://do i.org/ 10.1371/ journal.pcbi.10 10316 Editor: Lusha Zhu,Peking University, CHINA Received: January 22,2022 Accepted: June 18,2022 Published: August 4,2022 Copyright: ©2022 Dubey etal.Thisisanopen access article distributed under theterms ofthe Creative Commons Attribution License, which permits unrestricte duse,distribu tion,and reproduction inanymedium, provided theoriginal author andsource arecredited. Data Availabilit yStatement: Thesource code to produce theresults presented inthismanuscript is available onaGitHub repository athttps://github. com/rach00 12/happiness _RL. Funding: Theauthors received nospecific funding forthiswork. Competing interests :Theauthors have declared thatnocompeting interests exist.

2312.11671v2.pdf

Evaluating Language-Model Agents on Realistic Autonomous Tasks Megan Kinniment Lucas Jun Koba Sato Haoxing Du Brian Goodrich Max Hasin Lawrence Chan Luke Harold Miles Tao R. Lin Hjalmar Wijk Joel Burget Aaron Ho Elizabeth Barnes∗Paul Christiano† METR (Formerly ARC Evals) Abstract In this report, we explore the ability of language model agents to acquire resources, create copies of themselves, and adapt to novel challenges they encounter in the wild. We refer to this cluster of capabilities as “autonomous replication and adaptation” or ARA. We believe that systems capable of ARA could have wide- reaching and hard-to-anticipate consequences, and that measuring and forecasting ARA may be useful for informing measures around security, monitoring, and alignment. Additionally, once a system is capable of ARA, placing bounds on a system’s capabilities may become significantly more difficult. We construct four simple example agents that combine language models with tools that allow them to take actions in the world. We then evaluate these agents on 12 tasks relevant to ARA. We find that these language model agents can only complete the easiest tasks from this list, although they make some progress on the more challenging tasks. Unfortunately, these evaluations are not adequate to rule out the possibility that near-future agents will be capable of ARA. In particular, we do not think that these evaluations provide good assurance that the “next generation” of language models (e.g. 100x effective compute scaleup on existing models) will not yield agents capable of ARA, unless intermediate evaluations are performed during pretraining. Relatedly, we expect that fine-tuning of the existing models could produce substantially more competent agents, even if the fine-tuning is not directly targeted at ARA. 1 Introduction and motivation Large language models (LLMs) may cause significant real-world harm if they are used maliciously or pursue unintended goals. The extent of potential harms, and the necessary levels of caution, depend on models’ capabilities. Unfortunately, existing benchmarks often provide limited information about dangerous capabilities: risk depends on the behavior of AI systems in real-world environments, while benchmarks typically measure the performance of language models in short self-contained tasks like multiple choice tests or programming contests. ∗Corresponding author. Please direct correspondence to beth@evals.alignment.org. †Alignment Research Center.arXiv:2312.11671v2 [cs.CL] 4 Jan 2024

2310.11589.pdf

ELICITING HUMAN PREFERENCES WITH LANGUAGE MODELS Belinda Z. Li∗ MIT CSAIL bzl@mit.eduAlex Tamkin∗ Anthropic† atamkin@cs.stanford.eduNoah Goodman Stanford ndg@stanford.eduJacob Andreas MIT CSAIL jda@mit.edu ABSTRACT Language models (LMs) can be directed to perform target tasks by using labeled examples or natural language prompts. But selecting examples or writing prompts for can be challenging—especially in tasks that involve unusual edge cases, de- mand precise articulation of nebulous preferences, or require an accurate mental model of LM behavior. We propose to use LMs themselves to guide the task spec- ification process. In this paper, we introduce generative active task elicitation (GATE ): a learning framework in which models elicit and infer intended behavior through free-form, language-based interaction with users. We study GATE in three domains: email validation, content recommendation, and moral reasoning. In pre- registered experiments, we show that LMs prompted to perform GATE (e.g., by generating open-ended questions or synthesizing informative edge cases) elicit re- sponses that are often more informative than user-written prompts or labels. Users report that interactive task elicitation requires less effort than prompting or exam- ple labeling and surfaces novel considerations not initially anticipated by users. Our findings suggest that LM-driven elicitation can be a powerful tool for align- ing models to complex human preferences and values.1 1 I NTRODUCTION The complexity of human preferences makes them challenging to encode in machine learning sys- tems. Consider the problem of designing a recommendation system for songs or websites: first, system builders must develop a formal model of the potential factors influencing user preferences; second, users must describe their preferences in a format that a learning algorithm can use to make future recommendations. Each of these steps requires mental effort and continual refinement by users and system builders. Until recently, the dominant approach in machine learning has specified preferences using examples : users first label a dataset with examples of the desired model behavior, then train a machine learning model on this dataset. This strategy has seen widespread use across diverse tasks, including image classification and question answering (Krizhevsky et al., 2012; De- vlin et al., 2019). In more recent years, this paradigm has changed with the advent of instruction following methods (Brown et al., 2020a): by pre-training langauge models (LMs) on large-scale text corpora, it is possible to induce desired behaviors by conditioning only on natural language task specifications, in tasks as diverse as code generation and text summarization. However, this progress has also accentuated the challenges described above: complex behaviors require an increasing amount of prompt engineering ordataset design to overcome the imprecision of natural language and prevent models from misunderstanding or misgeneralizing from spurious features of prompts or examples. For example, a user who says they enjoy reading tennis articles could either be interested in the competitive tennis circuit or in improving their own serve. A few user-provided examples of tennis-related articles might fail to specify whether the user is interested in broader tennis content, such as tennis-themed satire. These challenges of task ambiguity (Finn et al., 2018; Tamkin et al., 2022a) loom large as models continue to be applied to more open-ended tasks and higher-stakes domains. ∗Equal contribution. Author order decided via coin flip. †Work performed while at Stanford University. 1Code is available at https://github.com/alextamkin/generative-elicitation 1arXiv:2310.11589v1 [cs.CL] 17 Oct 2023

2403.20222v1.pdf

Shallow Cross-Encoders for Low-Latency Retrieval Aleksandr V. Petrov, Sean MacAvaney, and Craig Macdonald University of Glasgow, Glasgow, UK a.petrov.1@research.gla.ac.uk {sean.macavaney;craig.macdonald }@glasgow.ac.uk Abstract. Transformer-based Cross-Encoders achieve state-of-the-art effectivness in text retrieval. However, Cross-Encoders based on large transformer models (such as BERT or T5) are computa- tionally expensive and allow for scoring only a small number of documents within a reasonably small latency window. However, keeping search latencies low is important for user satisfaction and energy usage. In this paper, we show that weaker shallow transformer models (i.e. transformers with a limited number of layers) actually perform better than full-scale models when constrained to these practical low-latency settings, since they can estimate the relevance of more documents in the same time budget. We further show that shallow transformers may benefit from the gen- eralised Binary Cross-Entropy (gBCE) training scheme, which has recently demonstrated success for recommendation tasks. Our experiments with TREC Deep Learning passage ranking querysets demonstrate significant improvements in shallow and full-scale models in low-latency scenarios. For example, when the latency limit is 25ms per query, MonoBERT-Large (a cross-encoder based on a full-scale BERT model) is only able to achieve NDCG@10 of 0.431 on TREC DL 2019, while TinyBERT-gBCE (a cross-encoder based on TinyBERT trained with gBCE) reaches NDCG@10 of 0.652, a +51% gain over MonoBERT-Large. We also show that shallow Cross-Encoders are effec- tive even when used without a GPU (e.g., with CPU inference, NDCG@10 decreases only by 3% compared to GPU inference with 50ms latency), which makes Cross-Encoders practical to run even without specialised hardware acceleration. 1 Introduction The introduction of the Transformer [35] neural network architecture, and especially pre-trained lan- guage models that use Transformers (such as BERT [7]), has been transformative for the IR field; for example, Nogueira et al. [27] improved MRR@10 on the MS-MARCO dev set by 31% with the help of a BERT-based model. Although there are a variety of ranking architectures used within IR (e.g., dense Bi-Encoders [14,18,40], sparse Bi-Encoders [9,22], and late interaction models [13,15]), the best results for document re-ranking are typically achieved with the help of Cross-Encoders [14] – a family of mod- els which encode both the query and the document simultaneously as a single textual input [41]. Aside from their high in-domain precision, Cross-Encoders tend to be more robust when generalising across retrieval tasks/domains [33]. Although Cross-Encoders can only practically be used as re-ranking models, limitations in their first-stage recall can be efficiently mitigated using pseudo-relevance feedback [23]. Further, Cross-Encoders can typically be fine-tuned from scratch (i.e., starting from the checkpoint of a foundational model, such as BERT). Despite these benefits, the application of Cross-Encoders in production retrieval systems is still limited. Cross-Encoders require a model inference for each query-document pair and, therefore, struggle with high computational complexity and high latency [24]. In real-world search systems, high latency negatively affects key performance metrics, such as the number of clicks, revenue, and user satisfaction [16, Ch. 5]. Further, high latencies tend to be correlated with higher energy usage, resulting in negative impacts on the climate [32]. The high computational complexity and resulting latency of Cross-Encoder models motivated re- searchers to investigate Bi-Encoder [14] models. These models separately encode the query and the docu- ment, and then estimate relevance score using an inexpensive operation over the encoded representations (e.g. cosine similarity [30] or the MaxSim operation [15]). By pre-computing the document representations offline and using a variety of approaches to accelerate retrieval [17], Bi-Encoders can achieve low retrieval latency. However, this comes at other costs. For instance, Bi-Encoders are markedly more complicated to train than Cross-Encoders, typically relying on knowledge distillation from other models (e.g., [19]), training data balancing (e.g., [12]), and/or hard negative mining (e.g., [40]). Further, Bi-Encoders must pre-encode all documents in the collection and keep the encoded versions of all documents in memory.arXiv:2403.20222v1 [cs.IR] 29 Mar 2024

2309.10400v3.pdf

Published as a conference paper at ICLR 2024 POSE: E FFICIENT CONTEXT WINDOW EXTENSION OF LLM S VIA POSITIONAL SKIP-WISE TRAINING Dawei Zhu∗♡♠Nan Yang♢Liang Wang♢Yifan Song♡♠Wenhao Wu♡♠ Furu Wei♢Sujian Li♡♠ ♡School of Computer Science, Peking University ♠National Key Laboratory for Multimedia Information Processing, Peking University ♢Microsoft Corporation https://github.com/dwzhu-pku/PoSE ABSTRACT Large Language Models (LLMs) are trained with a pre-defined context length, restricting their use in scenarios requiring long inputs. Previous efforts for adapting LLMs to a longer length usually requires fine-tuning with this target length ( Full- length fine-tuning), suffering intensive training cost. To decouple train length from target length for efficient context window extension, we propose Positional Skip-wis E(PoSE) training that smartly simulates long inputs using a fixed context window. This is achieved by first dividing the original context window into several chunks, then designing distinct skipping bias terms to manipulate the position indices of each chunk. These bias terms and the lengths of each chunk are altered for every training example, allowing the model to adapt to all positions within target length. Experimental results show that PoSE greatly reduces memory and time overhead compared with Full-length fine-tuning, with minimal impact on per- formance. Leveraging this advantage, we have successfully extended the LLaMA model to 128k tokens using a 2k training context window. Furthermore, we empir- ically confirm that PoSE is compatible with all RoPE-based LLMs and position interpolation strategies. Notably, our method can potentially support infinite length, limited only by memory usage in inference. With ongoing progress for efficient inference, we believe PoSE can further scale the context window beyond 128k. 1 I NTRODUCTION Large Language Models (LLMs) have revolutionized language modeling and demonstrated impres- sive abilities to perform various tasks (Brown et al., 2020). However, even with their remarkable capacity, these LLMs remain restricted by pre-defined context window sizes, suffering from notable performance decline when input tokens exceeds these limits. Nevertheless, numerous application scenarios demand extremely long input sequences, including long document summarization (Huang et al., 2021), in-context learning with numerous examples (Li et al., 2023), and long document retrieval (Zhou et al., 2022), etc. This naturally poses a significant challenge of context window extension : Extending the context window of a pre-trained LLM to accommodate longer sequences. Naively fine-tuning LLMs on inputs of target length for window extension has received limited success due to the large disruption introduced by new position indices (Chen et al., 2023a; Han et al., 2023). Addressing this, Position Interpolation (Chen et al., 2023a; kaiokendev, 2023; Peng et al., 2023) propose to down-scale the position indices to match the original window size, yielding improved results for context extension. However, these methods still rely on Full-length fine-tuning, i.e., fine- tuning with context of target length, which is memory and time-intensive due to the computational complexity that increases quadratically with input length. For example, Chen et al. (2023a) use 32 A100 GPUs to extend LLaMA models from 2k to 8k context, and 128 A100 GPUs for even larger context. These overhead has made it impossible to extend context window to extreme lengths. ∗Work done during Dawei’s internship at MSRA. Sujian Li is the corresponding author. 1arXiv:2309.10400v3 [cs.CL] 21 Feb 2024

2202.04728.pdf

Predicting Human Similarity Judgments Using Large Language Models Raja Marjieh1,*, Ilia Sucholutsky2,*, Theodore R. Sumers2, Nori Jacoby3, Thomas L. Griffiths1,2 1Department of Psychology, Princeton University 2Department of Computer Science, Princeton University 3Computational Auditory Perception Group, Max Planck Institute for Empirical Aesthetics {raja.marjieh, is2961, sumers, tomg }@princeton.edu; nori.jacoby@ae.mpg.de *equal contribution. Abstract Similarity judgments provide a well-established method for ac- cessing mental representations, with applications in psychol- ogy, neuroscience and machine learning. However, collecting similarity judgments can be prohibitively expensive for natu- ralistic datasets as the number of comparisons grows quadrati- cally in the number of stimuli. One way to tackle this problem is to construct approximation procedures that rely on more ac- cessible proxies for predicting similarity. Here we leverage recent advances in language models and online recruitment, proposing an efficient domain-general procedure for predicting human similarity judgments based on text descriptions. Intu- itively, similar stimuli are likely to evoke similar descriptions, allowing us to use description similarity to predict pairwise similarity judgments. Crucially, the number of descriptions required grows only linearly with the number of stimuli, dras- tically reducing the amount of data required. We test this pro- cedure on six datasets of naturalistic images and show that our models outperform previous approaches based on visual infor- mation. Keywords: similarity, perception, language models, represen- tations Introduction Mental representations serve as a substrate for a variety of cognitive tasks such as decision-making, communication and memory (Anderson, 1990). Understanding the structure of those representation is a core problem in cognitive science and is the subject of a large corpus of work in the psycho- logical literature (Shepard, 1980, 1987; Ghirlanda & Enquist, 2003; Battleday, Peterson, & Griffiths, 2020; Peterson, Ab- bott, & Griffiths, 2018; Jha, Peterson, & Griffiths, 2020; Caplette & Turk-Browne, 2022; Hebart, Zheng, Pereira, & Baker, 2020). One important example of this research is the development of the multi-dimensional scaling method (MDS) for uncover- ing the structure of mental representations based on similarity judgments (Shepard, 1980). Given a set of Nstimuli, MDS begins by collecting pairwise similarity judgments and aggre- gating them into a N×Nmatrix. Then, an iterative procedure finds an embedding that maps the stimuli into points in a psy- chological space such that their distance mirrors their simi- larity. Applying MDS to different datasets revealed highly in- terpretable organization of the stimuli (Shepard, 1980, 1987). Aside from psychology, similarity judgments play an impor- tant role in other disciplines such as neuroscience, e.g., in the method of representational similarity analysis (Kriegeskorte, Mur, & Bandettini, 2008), as well as in machine learning,e.g., as a way to regularize latent spaces so that they align with human representations and perception (Esling, Bitton, et al., 2018). Despite the success of these approaches, the quadratic in- crease of the number of pairwise comparisons as a function of the number of stimuli poses a serious limitation on their scalability. Indeed, even a relatively small dataset that con- tains∼102stimuli would require ∼104judgments for con- structing the full similarity matrix. This limitation calls for alternative procedures that allow for efficient approximation of human similarity judgments. Previous studies have pro- posed such a method in the visual modality by harnessing the latent representations from convolutional neural networks (CNNs) (Peterson et al., 2018; Jha et al., 2020). Such an approach, however, is domain-specific and could potentially miss important semantic dimensions that weigh on people’s judgments. To reduce this burden, we leverage the deep relationship between conceptual structure and language (Murphy, 2002) to use linguistic descriptions as a proxy for human seman- tic representations. Intuitively, stimuli that are judged to be highly similar are likely to evoke similar descriptions, allow- ing us to use description similarity to predict pairwise sim- ilarity judgments. This approach offers two key advantages over prior work: first, it is scalable . While pairwise similar- ity comparisons scale quadratically with the number of stim- uli (Shepard, 1980), text descriptions scale linearly. Second, it is domain-general : unlike CNN representations (Peterson et al., 2018), which are limited to visual stimuli, our proce- dure could be applied to any domain. Finally, we note that our approach leverages two distinct and important advances. First, text descriptions can be easily crowd-sourced via online recruitment platforms such as Ama- zon Mechanical Turk (AMT; https://www.mturk.com/ ) and are part of the common practice in modern machine learning pipelines (Parekh, Baldridge, Cer, Waters, & Yang, 2020). Second, modern language models (Speer, Chin, & Havasi, 2017; Devlin, Chang, Lee, & Toutanova, 2018) pro- vide rich latent representations of text. It is therefore natu- ral to ask: how far can we go in predicting human similarity judgments based on language alone? We explore this question on a collection of six datasets of naturalistic images for which the ground-truth similarity matrices are known (Peterson et al., 2018). Our explorationarXiv:2202.04728v1 [cs.LG] 9 Feb 2022