Stepwise Informativeness Search for Effective and Efficient LLM Reasoning

Siyuan Wang, Enda Zhao, Zhongyu Wei, Xiang Ren · 2025

Read Stepwise Informativeness Search for Effective and Efficient LLM Reasoning by Siyuan Wang, Enda Zhao, Zhongyu Wei, Xiang Ren (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-2502-15335.txt.

Description

Siyuan Wang, Enda Zhao, Zhongyu Wei and Xiang Ren sharpen step-by-step reasoning in language models, in work from 2025. Long chains of thought tend to drift: later steps forget early facts and repeat what is already known. Their stepwise informativeness search is a tree search run at inference time, steered by two heuristics.

One favours steps that attend to neglected earlier steps, and a self-grounding prompt makes the model cite its premises before each deduction. The other favours steps whose conclusions are new. On four reasoning datasets the method raised accuracy while producing shorter rationales with fewer errors and less redundancy than standard decoding.

Catalog record

CREATORSiyuan Wang, Enda Zhao, Zhongyu Wei, Xiang Ren
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-15335

Browse the full library · About the Tower of Babel