IAAISTARS
Stabilizing Recurrent Dynamics for Test-Time Scalable Latent Reasoning in Looped Language Models
Addresses the peak-then-collapse behavior of looped language models under deeper recurrence by constraining latent states to approach asymptotically stable fixed points, using efficient Jacobian spectral-radius regularization with random loop sampling.
Inside the method
[ Block ] × r with Jacobian spectral-radius penalty
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Core recurrence
- Depth control
- Test-time scalable recurrence
- KV / state strategy
- See paper
Reading note
Gives reliable test-time scaling on arithmetic tasks and on harder math reduces degradation at depth while also improving peak accuracy.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{stars2026,
title = {Stabilizing Recurrent Dynamics for Test-Time Scalable Latent Reasoning in Looped Language Models},
author = {Xiao-Wen Yang and Ziyu Han and Xi-Hua Zhang and Wen-Da Wei and Jie-Jing Shao and Lan-Zhe Guo and Yu-Feng Li},
year = {2026},
eprint = {2605.26733},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2605.26733}
}