When Do LLMs Reason? A Dynamical Systems View via Entropy Phase Transitions

Wei Xia, Haoqing Wang, Yehui Tang, Zhi-Hong Deng · 2026

Read When Do LLMs Reason? A Dynamical Systems View via Entropy Phase Transitions by Wei Xia, Haoqing Wang, Yehui Tang, Zhi-Hong Deng (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-2605-22873.txt.

Description

Xia and three colleagues ask when chain-of-thought reasoning actually helps language models, and route decoding by early entropy dynamics. Tasks with large reasoning gains tend to show falling early entropy, while low-gain tasks oscillate or rise. Their EDRM framework summarizes those early trajectories and chooses among direct, standard, and chain-of-thought decoding. Across 15 benchmarks and 7 models it saved 16 to 55 percent of tokens at comparable accuracy.

Catalog record

CREATORWei Xia, Haoqing Wang, Yehui Tang, Zhi-Hong Deng
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE102854
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.22873

Browse the full library · About the Tower of Babel