text
stringlengths 62
4.5k
| arxiv-translator
stringlengths 51
2.86k
|
---|---|
A more conservative API provider may remove access to the combination of logit bias and logprobs entirely. Indeed, after disclosing our attack to OpenAI, they removed the ability for logit bias to impact the top logprobs--thus preventing the attacks from the prior sections. To exploit situations such at this, we further develop several logprob-free attacks that recover the complete logit vector by performing binary search on the logit bias vector, albeit at increased cost. | λ³΄λ€ λ³΄μμ μΈ API 곡κΈμλ λ‘μ§ νΈν₯κ³Ό λ‘κ·Ένλ‘λΈμ μ‘°ν©μ λν μ‘μΈμ€λ₯Ό μμ ν μ κ±°ν μ μλ€. μ€μ λ‘ μ€νAIμ μ°λ¦¬μ 곡격μ 곡κ°ν ν, κ·Έλ€μ λ‘μ§ νΈν₯μ΄ μμ λ‘κ·Ένλ‘λΈμ μν₯μ λ―ΈμΉλ λ₯λ ₯μ μ κ±°νμ¬ μ΄μ μΉμ
μμ 곡격μ λ°©μ§νλ€. μ΄λ¬ν μν©μ νμ©νκΈ° μν΄ λΉμ© μ¦κ°μλ λΆκ΅¬νκ³ λ‘μ§ λ°μ΄μ΄μ€ 벑ν°μ λν΄ μ΄μ§ κ²μμ μννμ¬ μμ ν λ‘μ§ λ²‘ν°λ₯Ό 볡ꡬνλ μ¬λ¬ λ‘κ·Ένλ‘λΈ μλ 곡격μ μΆκ°λ‘ κ°λ°νλ€. |
3. 30/30 workout: 30 seconds of intense exercise followed by 30 seconds of rest for several rounds. | 3. 30/30 μ΄λ: 30μ΄μ κ²©λ ¬ν μ΄λ ν 30μ΄μ ν΄μμ μ¬λ¬ λΌμ΄λ λμ μννλ€. |
Footnote 6: There is evidence from Grosse et al. (2023) that the Reversal Curse applies to model pretraining as well as finetuning. For cost reasons, we tested finetuning rather than pretraining. | κ°μ£Ό 6: Grosse et al.(2023)μ μ¦κ±°κ° μλλ°, μμ μ£Όλ λͺ¨λΈ μ¬μ νλ ¨λΏλ§ μλλΌ λ―ΈμΈ μ‘°μ μλ μ μ©λλ€λ κ²μ΄λ€. λΉμ©μ μΈ μ΄μ λ‘ μ¬μ νλ ¨λ³΄λ€λ λ―ΈμΈ μ‘°μ μ ν
μ€νΈνμ΅λλ€. |