IAAIMonitoring Latent Reasoning
Monitoring latent reasoning in looped language models
An interpretability-based inspection protocol with five lenses applied to eight looped language models from four families, asking whether iterations actually improve capability and whether models stay stable under prompt perturbation.
Inside the method
Carry loop (prompt only in state) ‖ Re-read loop (prompt re-supplied each iteration)
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Full-stack recurrence
- Depth control
- Analysis of released models
- KV / state strategy
- See paper
Reading note
The key axis is whether the prompt lives only in the looped state (Ouro, LoopUS) or is re-read every iteration (Parcae, Huginn, Rec-OLMo). Findings are framed around safety and efficient use.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{monitoringloopedlms2026,
title = {Monitoring latent reasoning in looped language models},
author = {Pawel Olszowiec and Paulina Tomaszewska and Maciej Żelaszczyk and Michal Byra and Neo Christopher Chung},
year = {2026},
eprint = {2610.18127},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.18127}
}