ThoughtProbe: Classifier-Guided LLM Thought Space Exploration via Probing Representations

Zijian Wang, Chang Xu · 2025

Read ThoughtProbe: Classifier-Guided LLM Thought Space Exploration via Probing Representations by Zijian Wang, Chang Xu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-27355.txt.

Description

Zijian Wang and Chang Xu give language models a smarter way to shop among their own thoughts, in a 2025 paper. A model can branch its reasoning into a tree of paths, then pick the best ending. Blind branching wastes effort on weak paths. Their framework, ThoughtProbe, scores each branch early using the model's inner signals.

A small classifier reads those hidden signals and ranks the branches. Compute goes to the high scorers first. When the tree is done, all branch answers form a pool. A pooling step then weighs each answer by the scores of the branches behind it. The best-backed answer wins. On several arithmetic reasoning tests, the method found and picked good paths far more often than plain search.

Catalog record

CREATORZijian Wang, Chang Xu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE50 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-27355

Browse the full library · About the Tower of Babel