flâneur

Kaiyue Wen on X: "Can LLMs generate novel research ideas? We propose Priced Guidance to quantify this via compression! We lower bound P(LLM forecasts a future paper’s idea X). Instead of sampling ideas until X is matched, the LLM plays “20 Questions” to recover X, charging each hint by surprisal. https://t.co/QKE1byIV2Y" / X

x.com · 294 words · saved by 1 readers

Can LLMs generate novel research ideas? We propose Priced Guidance to quantify this via compression! We lower bound P(LLM forecasts a future paper’s idea X). Instead of sampling ideas until X is matched, the LLM plays “20 Questions” to recover X, charging each hint by surprisal.

(2/9) How it works: the LLM (the generator) asks multiple-choice questions, e.g. "what is the topic?", with a probability p on each answer; a guide LLM that knows the paper picks one, charging −log₂ p bits. E.g., Fable 5 rediscovers SkillNet in 19 questions and 64 bits. (3/9) We prove that if a generator recovers the idea with K bits of guidance, the guide's answers can be compressed to K + O(1) bits and the generator can generate the idea with no guidance w.p. ≥ 2^−K. Similar to language modelling, better compression => better generation! (4/9) We test frontier LLMs on 87 of the recent…

saved by

related reading