IAAIDissecting HRM
Dissecting Hierarchical Reasoning Models: A Mechanistic Study
A mechanistic study of the Hierarchical Reasoning Model on Sudoku, Maze and ARC-AGI-2 using baselines with and without recurrence, state interventions, probes and sparse autoencoders.
Inside the method
[ L-module × n → H-module ] × m on a puzzle state
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Core recurrence
- Depth control
- Analysis
- KV / state strategy
- Not applicable (puzzle solver)
Reading note
Recurrent models beat one-pass baselines, but a single-state recurrent Transformer is comparable to HRM. HRM behaves like constraint-aware iterative refinement without a compact, causally important feature set.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{dissectinghrm2026,
title = {Dissecting Hierarchical Reasoning Models: A Mechanistic Study},
author = {Leo Raphael Rodrigues and Jian Kang},
year = {2026},
eprint = {2609.22197},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.22197}
}