Percy Liang 提出 priced guidance 评估 Astra 能否预测新论文
精选理由
Percy Liang 提了个挺新的评估思路：让带答案的 guide 模型和裸模型对话，按传输比特数记账，算出模型预测新论文的概率下界。
Percy Liang 讨论如何估算 Astra 生成截止日期之后某篇新论文的概率，直接朴素采样会因尾概率过小而不可行。priced guidance 的核心做法是让模型 Astra 与一个已读论文的 guide LLM 对话，通过精确统计 guide 传给模型的比特数来给出概率下界。这种思路也可以作为面向前沿模型的新评估方式。
原文 · Percy Liang
What is the probability that Astra today can generate/forecast a particular new paper past its cutoff date? Naive sampling is intractable (tail probability is astronomically small).
The key idea of priced guidance is that you can lower bound this probability via a dialogue between the model (Astra) and a guide LLM (Astra with the paper), where we can precisely account for the bits of information sent from the guide to the model. One can also can think of this as a new eval for frontier models.
