Hierarchical Reasoning Model Reasoning, the process of devising and executing complex goal-oriented action sequences, remains a critical challenge in AI. Current large language models (LLMs) primarily employ Chain-of-Thought (CoT...

The original HRM paper by Wang et al can be found here:
arxiv.org/abs/2506.21734

GitHub - LongDangHoang/HRM_RL_Agent: HRM Agent repo HRM Agent repo. Contribute to LongDangHoang/HRM_RL_Agent development by creating an account on GitHub.

Our paper is simply a proof-of-concept and our model is the simplest possible RL agent version of HRM. We're looking forward to working with more sophisticated recurrent reasoning models like HRM and TRM on more complex problems.

Code:
github.com/LongDangHoan...

Divergence from initial latent state during recurrent processing. When the latent state from the previous environmental time step is carried forward, the initial latent state is more similar to the final latent state.

In addition, by analyzing the divergence of the latent state we found evidence that the resulting plans (paths) are more consistent over time.

In dynamic environments, it's important for plan continuity and efficiency to reuse computation from previous environment time-steps. We found that the recurrent process in the HRM-Agent model converges more quickly when the latent state is copied across from previous time-steps.

The Hidden Drivers of HRM's Performance on ARC-AGI We scored on hidden tasks, ran ablations, and found that performance from the Hierarchical Reasoning Model comes from an unexpected source

The ARC-prize team came to similar conclusions about the importance of recurrence during inference:
arcprize.org/blog/hrm-ana...

Less is More: Recursive Reasoning with Tiny Networks Hierarchical Reasoning Model (HRM) is a novel approach using two small neural networks recursing at different frequencies. This biologically inspired method beats Large Language models (LLMs) on hard ...

Why is HRM so efficient at reasoning tasks? Check out "Less is More: Recursive Reasoning with Tiny Networks" by Alexia Jolicoeur-Martineau. Her simplified model explores the recurrent concept introduced in HRM which seems to be responsible for much of the performance.
arxiv.org/abs/2510.04871

Plot showing fraction of validation episodes in which the agent reached the goal, from 5 runs with carry Z condition and 5 runs with reset Z condition.

We found that the HRM-Agent can learn to navigate in dynamic and uncertain maze environments, with doors which open and close randomly.

Dynamic maze environment from the paper, a screenshot from the nethack learning environment (NLE)

The Hierarchical Reasoning Model (HRM) has impressive reasoning abilities given its small size, but has only been applied to supervised, static, fully-observable problems.

We wanted to see if we could train a HRM to navigate in a maze using only reinforcement learning.

Long H Dang, David Rawlinson: HRM-Agent: Training a recurrent reasoning model in dynamic environments using reinforcement learning https://arxiv.org/abs/2510.22832 https://arxiv.org/pdf/2510.22832 https://arxiv.org/html/2510.22832

AI Revolution, Meet the Credibility Revolution Introducing Workhelix

geekway.substack.com/p/ai-revolut... #causalsky Blog post on using causal inference to understand the ROI of generative AI

I couldn't find Conf_CLeaR or any existing post on BlueSky...

CLeaR-Conference on Causal Learning and Reasoning on X: "Interested in Causality and its many applications? Don't forget to register for #CLeaR2025, the 4th Conference on Causal Learning and Reasoning, conference May 7-9 in Lausanne, Switzerland! ✅Register ASAP by Mar 31 to enjoy the early-bird discount: https://t.co/lOiIQuPvYA https://t.co/iFAgOBCqkn" / X Interested in Causality and its many applications? Don't forget to register for #CLeaR2025, the 4th Conference on Causal Learning and Reasoning, conference May 7-9 in Lausanne, Switzerland! ✅Register ASAP by Mar 31 to enjoy the early-bird discount: https://t.co/lOiIQuPvYA https://t.co/iFAgOBCqkn

x.com/Conf_CLeaR/s...

Looking forward to seeing you at Spatiotemporal Causal Analysis (#STCausal2025), stcausal2025.spatial-causal.org with @grantdmckenzie.bsky.social and Cecile de Bezenac!

As reported, the entire staff of 538 was laid off this morning. This is a severe blow to political data journalism, and I feel for my colleagues. Readers note: As we were instructed not to publish any new content, all planned updates to polls data and averages are canceled indefinitely. Huge loss :(

Dave Lagnado is looking for a post doc to work on causal inference! The BR-UK team is doing some very cool stuff, so if you're currently looking for a job, check this out: www.ucl.ac.uk/work-at-ucl/...

This is a brilliantly clear thread from Julia about "why causal inference from observational data is difficult". This, combined with my ongoing re-read of The Book of Why, is finally building an intuition about causality into my feeble human brain

Is this is fabled Kebab Collider that I mention in some of my lectures? Does look like filtering out of [far+bad] restaurants. Given spatial confounding, stronger pattern than I would have expected. Wonder what a simulation on same urban network would look like. ht @erikwestlund.bsky.social

"In short, QCA often finds complexity where none exists."

This is a great new ETP editorial by my coauthor Mikko Rönkkö, Markku Maula, and Karl Wennberg, illustrating the massive false positives problem in Qualitative Comparative Analysis (QCA). doi.org/10.1177/1042...

Matching, missing data, a quasi-experiment, and causal inference--Oh my! | A. Solomon Kurz I'm finally dipping my does into causal inference for quasi-experiments, and my first use case has missing data. In this post we practice propensity score matching with multiply-imputed data sets, and...

New blog up: solomonkurz.netlify.app/blog/2025-02...

This time I dip my toes into causal inference for quasi-experiments using matching methods, and my use case has missing data complications. Many thanks to @dingdingpeng.the100.ci and
@noahgreifer.bsky.social
for their peer review! #RStats

Can we make this the official way to draw the set of unobserved confounders...?

That feeling when you've carefully embedded the figures in the right place and the journal has a template which insists they all have to go at the end 😭

"Causal Discovery in Python" - Lizzie Silver (Pycon AU 2024) YouTube video by PyCon AU

Check out my talk: Causal Discovery in Python www.youtube.com/watch?v=M2lL...

@dagophile.bsky.social drops some powerful bombs in this episode.

One of the most important voices at the intersection of causality, philosophy and dynamical systems.

#CausalSky

Every time someone makes a causal claim based on a single plot, a kitten dies .or "The Truth is Simple" The complexity of the description of a dog resting in a position that isn't clearly captured by common-sense terms forces us to use more words to accurately convey the scene 1/ 👇🏼 #CausalSky

Every time someone makes a causal claim based on a single plot, a kitten dies

.or "The Truth is Simple"

The complexity of the description of a dog resting in a position that isn't clearly captured by common-sense terms forces us to use more words to accurately convey the scene

1/ 👇🏼

#CausalSky

DEADLINE EXTENDED🚨 Revolutionize #healthcare with data, causal inference, and machine learning! Join the Causal Risk Prediction in Medicine Datathon at the I2DB Symposium on Feb. 12, 2025. Register your team by Dec. 13.

Learn more ➡️ i2db.wustl.edu/calendar_eve...

#WashUMed #Datathon #RegisterToday

Advances in Difference-in-differences Methods for Policy... : Epidemiology s literature has revealed that DiD estimators may exhibit bias when heterogeneous treatment effects, a common consequence of staggered policy implementation, are present. To deepen our understanding o...

journals.lww.com/epidem/fullt...

#causalsky #causalinference #episky #stats #statsky

Sharpening causal reasoning in applied ethnographic research
doi.org/10.1080/0018...

Figure 1. A Stepwise Illustration of Ethnographic Research Conceptualization and Design to Strengthen Causal Inference.

My personal view (which seems to align with several of the authors in the table) is that if we are going to have both terms, then gaining an understanding of an interpretable model does not require fitting an additional model, whereas explainable solutions rely on an additional explanation model.

Overview of definitions for interpretability and explainability for different papers. The gist: the authors have different definitions.

What's the difference between explainability and interpretability? Does the machine learning community have an agreed-upon definition?

No.

There's a great overview, which is from this paper: arxiv.org/abs/2211.08943

My take: I prefer interpretability since the term explainability is too strong.