hierarchical reasoning model
In plain wordsA small, brain-inspired AI model that solves puzzles with two linked parts — a slow planner and a fast worker — instead of writing out its reasoning in words.
A small recurrent network for puzzle-like reasoning that couples two modules ticking at different rates, loosely modeled on the brain's layered timescales: a slow module that revises an overall plan only occasionally, and a fast module that runs many quick refinement steps beneath it, converging before each slow update. All the reasoning happens inside its hidden states during one forward pass, with no written chain of thoughtchain of thought, n.: Getting an AI to write out its reasoning step by step before giving an answer, the way a student shows their working. This often makes answers to math and logic problems more accurate., and it was trained from scratch on task examples alone, with no pretraining or chain-of-thought supervision. Its authors reported that a 27-million-parameter version, given about 1,000 examples per task, solved extremely hard Sudoku almost perfectly, found optimal routes through large mazes and outscored far bigger language models on ARC-AGIARC-AGI, prop. n.: Puzzles made of colored grids — and, since 2026, small video-game-like worlds — that people solve easily but AI long found hard, designed to test whether a machine can learn a new skill from just a few examples.. Later analyses, including one by the ARC Prize organizers, attributed much of its performance to iterative refinement and data augmentation rather than to the hierarchy itself.
The hierarchical reasoning model solved the maze in one forward pass, without a single reasoning token.
Origin
Introduced by Guan Wang, Jin Li, Yuhao Sun, Xing Chen, Changling Liu, Yue Wu, Meng Lu, Sen Song and Yasin Abbasi Yadkori (Sapient Intelligence, with Tsinghua University collaborators) in a paper posted to arXiv on 26 June 2025, drawing on the brain's hierarchical, multi-timescale processing.
Related
- tiny recursive modeltiny recursive model, n.: A very small AI model that solves hard puzzles by going over its own answer again and again, improving it each time, instead of relying on sheer size.
- latent reasoninglatent reasoning, n.: Getting an AI to do its step-by-step thinking inside its internal numbers rather than writing it out in words — potentially faster and more flexible, but much harder for people to inspect.
- recurrent neural networkrecurrent neural network, n.: A neural network that reads a sequence, such as a sentence or an audio clip, one piece at a time, carrying a running summary forward so that what came earlier can shape how later pieces are understood.
- ARC-AGIARC-AGI, prop. n.: Puzzles made of colored grids — and, since 2026, small video-game-like worlds — that people solve easily but AI long found hard, designed to test whether a machine can learn a new skill from just a few examples.
- chain of thoughtchain of thought, n.: Getting an AI to write out its reasoning step by step before giving an answer, the way a student shows their working. This often makes answers to math and logic problems more accurate.
- System 1 and System 2System 1 and System 2, phr.: A popular way of describing two modes of thinking: fast, automatic and intuitive (System 1) versus slow, effortful and deliberate (System 2). AI researchers borrow it to contrast instant answers with step-by-step reasoning.
Sources
- Wang, G., Li, J., Sun, Y., Chen, X., Liu, C., Wu, Y., Lu, M., Song, S. & Abbasi Yadkori, Y. (2025). Hierarchical Reasoning Model. arXiv:2506.21734. arxiv.org/abs/2506.21734
Mentioned in
tiny recursive modeltiny recursive model, n.: A very small AI model that solves hard puzzles by going over its own answer again and again, improving it each time, instead of relying on sheer size.