IAAIMechanistic Analysis of Looped LMs
A Mechanistic Analysis of Looped Reasoning Language Models
Analyzes latent states of looped reasoning language models and finds that each layer in the cycle often converges to its own fixed point, so the recurrent block follows a consistent cyclic trajectory and repeats feedforward-like stages of inference at every iteration.
Inside the method
[ Layer 1 → … → Layer k ] × R → cyclic fixed points
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Core recurrence
- Depth control
- Analysis
- KV / state strategy
- See paper
Reading note
Attention-head behavior stabilizes as fixed points are reached. Recurrent block size, input injection and normalization shape whether these cyclic fixed points emerge.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{loopedreasoningmechanistic2026,
title = {A Mechanistic Analysis of Looped Reasoning Language Models},
author = {Hugh Blayney and Álvaro Arroyo and Johan Obando-Ceron and Pablo Samuel Castro and Aaron Courville and Michael M. Bronstein and Xiaowen Dong},
year = {2026},
eprint = {2604.11791},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2604.11791}
}