自己回帰モデルのLLMは必ず誤る

Yann LeCun

@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基素.icon

https://pbs.twimg.com/media/FsLg2grXoAAcSAV.jpg

@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.

https://drive.google.com/file/d/1BU5bV3X5w65DwSMapKcsr0ZvrMRU_Nbi/view?usp=share_link

@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.