How can we build AI that can solve reasoning puzzles? A recent paper, "Hierarchical Reasoning Model," shocked the AI community with promising results on Sudoku, maze puzzles, and ARC-AGI benchmarks. This video provides an overview of the Hierarchical Reasoning Model.

00:00 Reasoning tasks
00:22 Hierarchical Reasoning Models' results
01:07 Problem setup
02:00 Transformer
02:37 Chian-of-thought reasoning
03:14 Recurrent models
04:31 HRM - Architecture
06:12 HRM - Gradient approximation
07:48 Specialized vs general models

References:
- Hierarchical Reasoning Model: https://arxiv.org/abs/2506.21734
- End-to-end Algorithm Synthesis with Recurrent Networks: Logical Extrapolation Without Overthinking https://arxiv.org/abs/2202.05826
- Scaling up test-time compute with latent reasoning: A recurrent depth approach: https://arxiv.org/abs/2502.05171
- Looped Transformers are Better at Learning Learning Algorithms, https://arxiv.org/abs/2311.12424
- Looped Transformers as Programmable Computers, https://arxiv.org/abs/2301.13196

Video made with Manim: https://www.manim.community/

Jia-Bin Huang

A hierarchical reasoning model (HRM) with only 27 million parameters outperforms large language models like DeepSeek R1 and GPT-o3 on inductive reasoning benchmarks including Sudoku and maze navigation. The model uses a two-level recurrent architecture: low-level modules handle fast, detailed computations while high-level modules build abstract representations over longer time horizons. Key techniques include input injection to maintain problem context, fixed-point gradient approximation to reduce memory usage, deep supervision across segments, and adaptive computation that stops recurrence early when a solution is found. Remarkably, HRM achieves near-perfect accuracy on challenging Sudoku problems after training on just 1,000 examples.

The Weirdly Small AI That Cracks Reasoning Puzzles [HRM]