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

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...

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

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

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

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.

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

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 😭