Stabilizing Recurrent Dynamics for Test-Time Scalable Latent Reasoning in Looped Language Models

Xiao-Wen Yang, Zi-Yu Han, Xi-Hua Zhang, Wen-Da Wei, Jie-Jing Shao, Lan-Zhe Guo, Yu-Feng Li · 2026

Read Stabilizing Recurrent Dynamics for Test-Time Scalable Latent Reasoning in Looped Language Models by Xiao-Wen Yang, Zi-Yu Han, Xi-Hua Zhang, Wen-Da Wei, Jie-Jing Shao, Lan-Zhe Guo, Yu-Feng Li (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-2605-26733.txt.

Description

Yang and six colleagues address looped language models, which reason by iterating a single layer but often peak and then collapse as iteration depth grows. Their STARS training framework steers latent states toward asymptotically stable fixed points by regularizing the Jacobian spectral radius, a measure of how errors amplify across steps. On arithmetic and mathematical reasoning tasks, STARS delivers reliable test-time scaling, curbs late-depth degradation, and lifts peak performance.

Catalog record

CREATORXiao-Wen Yang, Zi-Yu Han, Xi-Hua Zhang, Wen-Da Wei, Jie-Jing Shao, Lan-Zhe Guo, Yu-Feng Li
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43044
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.26733

Browse the full library · About the Tower of Babel