自己回帰モデルのLLMは必ず誤る
> @ylecun: I have claimed that Auto-Regressive LLMs are exponentially diverging diffusion processes.
> Here is the argument:
> Let e be the probability that any generated token exits the tree of "correct" answers.正しい答えのtreeから外れる確率
> Then the probability that an answer of length n is correct is (1-e)^n
長さnの回答が正しい確率は(1-e)^n
つまり、nが増えるたびにものすごい勢いで正しい確率が減る
> 1/
>
> @ylecun: Errors accumulate.
> The proba of correctness decreases exponentially.
> One can mitigate the problem by making e smaller (through training) but one simply cannot eliminate the problem entirely.trainingを通じてeを小さくすることで問題を小さくできるが解決はできない
> A solution would require to make LLMs non auto-regressive while preserving their fluency.
LLMを流暢さを保持しながら自己回帰をやめることが必要
> @ylecun: The full slide deck is here.
> This was my introductory position statement to the philosophical debate
> “Do large language models need sensory grounding for meaning and understanding?”
> Which took place at NYU Friday evening.
>
> @ylecun: I should add that things like RLHF may reduce e but do not change the fact that token production is auto-regressive and subject to exponential divergence.
関連