aiqbench analysis
This mechanistic study reveals that Hierarchical Reasoning Models (HRM) function through constraint-aware iterative refinement rather than relying on a compact set of causally critical features. The analysis demonstrates that while recurrent architectures outperform one-pass baselines, top-ranked sparse autoencoder features do not offer stable advantages over random subsets during large-scale ablations. These findings underscore the necessity for developing interpretability techniques specifically tailored to latent-space, recursive reasoning models in AI.
Dissecting Hierarchical Reasoning Models: A Mechanistic Study
arXiv:2609.22197v1 Announce Type: new Abstract: We study Hierarchical Reasoning Model (HRM), a representative hierarchical Transformer-based latent reasoning model with many variants, on Sudoku, Maze, and ARC-AGI-2. We mechanistically understand how HRM reasons and what information it encodes. Our analyses compare HRM against Transformer baselines with and without recurrent modules, apply causal interventions on recurrent states, and utilize linear probes against random-direction ablations, as well as sparse autoencoders with feature ablations. Our results reveal several key findings: recurrent models outperform one-pass baselines, while single-state recurrent Transformers are comparable to HRM. State interventions further show that the causal contributions of the high- and low-level states vary across task-specific checkpoints and inference stages. Selected task variables are linearly decodable from the recurrent states in HRM, yet ablating probe directions produce effects comparable to random controls. SAE ablations yield larger behavioral changes than probe-direction ablations. However, top-ranked SAE features show no stable advantage over size-matched random subsets at larger ablation sizes or across tasks; the same pattern persists in a Sudoku control with within-step BPTT. Together, we characterize that HRM is essentially implementing constraint-aware iterative refinement on a puzzle-specific solution state, in which the functional contributions of components at different levels vary without relying on a compact, causally important feature set. These results highlight the necessity of studying the different working mechanisms and the importance of developing mechanistic interpretability techniques better suited for latent-space, recursive reasoning models.
Abstract:We study Hierarchical Reasoning Model (HRM), a representative hierarchical Transformer-based latent reasoning model with many variants, on Sudoku, Maze, and ARC-AGI-2. We mechanistically understand how HRM reasons and what information it encodes. Our analyses compare HRM against Transformer baselines with and without recurrent modules, apply causal interventions on recurrent states, and utilize linear probes against random-direction ablations, as well as sparse autoencoders with feature ablations. Our results reveal several key findings: recurrent models outperform one-pass baselines, while single-state recurrent Transformers are comparable to HRM. State interventions further show that the causal contributions of the high- and low-level states vary across task-specific checkpoints and inference stages. Selected task variables are linearly decodable from the recurrent states in HRM, yet ablating probe directions produce effects comparable to random controls. SAE ablations yield larger behavioral changes than probe-direction ablations. However, top-ranked SAE features show no stable advantage over size-matched random subsets at larger ablation sizes or across tasks; the same pattern persists in a Sudoku control with within-step BPTT. Together, we characterize that HRM is essentially implementing constraint-aware iterative refinement on a puzzle-specific solution state, in which the functional contributions of components at different levels vary without relying on a compact, causally important feature set. These results highlight the necessity of studying the different working mechanisms and the importance of developing mechanistic interpretability techniques better suited for latent-space, recursive reasoning models.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2609.22197 [cs.LG]
(or arXiv:2609.22197v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2609.22197 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 323 words
Related coverage
PIRAMID: Progress and Plans
LessWrong · Jul 27, 2026
Introducing PIRAMID: Physics-Informed Research for Ambitious Mechanistic Interpretability
LessWrong · Jul 25, 2026
Probing the Origins of Reasoning Performance: Representational Quality for Mathematical Problem-Solving in RL vs. SFT Fine-Tuned Models
arXiv · Jul 31, 2026
Generalization and infinite width
LessWrong · Aug 1, 2026