IAAIWhat Limits Recursive Reasoning
What Limits Recursive Reasoning Models: Optimization, Architecture and Test-Time Scaling
Puts HRM, TRM and URM-style recursive reasoners into one experimental pipeline across six algorithmic domains to separate the effects of architecture, gradient propagation and training procedure. A simple recipe emerges: an intermediate gradient horizon, large physical batches and controlled recurrent-state updates, with no explicit hierarchy needed.
Inside the method
[ Small shared block ] × N refining a latent state
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Core recurrence
- Depth control
- Many recurrent steps with an intermediate gradient horizon
- KV / state strategy
- Not applicable (non-autoregressive solver); see paper
Reading note
The resulting 13.6M-parameter model raises Arithmetic OOD accuracy to 71.2% from 36.2%, reaches 98.41% on Sudoku and 59.5% pass@2 on ARC-AGI-1.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{recursivereasoninglimits2026,
title = {What Limits Recursive Reasoning Models: Optimization, Architecture and Test-Time Scaling},
author = {Yuliana Shakhvalieva and Dmitrii Kharchev and Viacheslav Bezrukov and Inessa Fedorova and Dmitry Bocharov and Ivan Oseledets and Valerii Ternovskii},
year = {2026},
eprint = {2609.39967},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.39967}
}