Clear : Contrasting Textual Feedback with Experts and Amateurs for Reasoning

Andrew Rufail, Daniel Kim, Sean O’Brien, Kevin Zhu · 2025

Read Clear : Contrasting Textual Feedback with Experts and Amateurs for Reasoning by Andrew Rufail, Daniel Kim, Sean O’Brien, Kevin Zhu (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-2504-07116.txt.

Description

Andrew Rufail, Daniel Kim, Sean O'Brien, and Kevin Zhu introduce CLEAR, a reasoning method that improves a language model's answer by contrasting feedback from two other models. A large expert model and a small amateur model each critique the first draft. Their comments are set against each other to distill sharper, refined feedback.

That feedback is then applied to revise the answer, and the loop can repeat. Across several hard reasoning tasks, CLEAR beat leading methods. Story outlines became up to about a fifth more interesting, constrained writing covered more required points, math accuracy rose, and toxic output fell by up to about a fifth. The authors credit the contrast: the amateur's misses help reveal what the expert's advice really adds.

Catalog record

CREATORAndrew Rufail, Daniel Kim, Sean O’Brien, Kevin Zhu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-07116

Browse the full library · About the Tower of Babel