SLAT: Segment-Level Adaptive Trimming for Efficient CoT Reasoning

Jian Yao, Xiongcai Luo, Ran Cheng, Kay Chen Tan · 2026

Read SLAT: Segment-Level Adaptive Trimming for Efficient CoT Reasoning by Jian Yao, Xiongcai Luo, Ran Cheng, Kay Chen Tan (2026) 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-2605-30832.txt.

Description

Yao and colleagues address excess thought in reasoning models, where thought chains carry built-in repeats and high compute cost without lifting rightness. Old tactics often rely on token-uniform length fines, giving coarse pressure toward shorter outputs that can suppress useful thought. The authors show waste sits in high-odds segments with low marginal utility. They propose SLAT, a reinforcement learning framework selectively suppressing redundant segments based on this criterion. On standard benchmarks, SLAT establishes a superior accuracy-efficiency frontier, reducing thought length by half while maintaining competitive accuracy.

Catalog record

CREATORJian Yao, Xiongcai Luo, Ran Cheng, Kay Chen Tan
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE56135
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.30832

Browse the full library · About the Tower of Babel