You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
We introduce Reprompting, an iterative sampling algorithm that searches forthe Chain-of-Thought (CoT) recipes for a given task without human intervention.Through Gibbs sampling, we infer CoT recipes that work consistently well for aset of training samples. Our method iteratively samples new recipes usingpreviously sampled solutions as parent prompts to solve other trainingproblems. On five Big-Bench Hard tasks that require multi-step reasoning,Reprompting achieves consistently better performance than the zero-shot,few-shot, and human-written CoT baselines. Reprompting can also facilitatetransfer of knowledge from a stronger model to a weaker model leading tosubstantially improved performance of the weaker model. Overall, Repromptingbrings up to +17 point improvements over the previous state-of-the-art methodthat uses human-written CoT prompts.
AkihikoWatanabe
changed the title
あ
Reprompting: Automated Chain-of-Thought Prompt Inference Through Gibbs
Sampling, Weijia Xu+, N/A, arXiv'23
May 22, 2023
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)
The text was updated successfully, but these errors were encountered: