HRM

Why we included this project

For researchers who want to see a different approach to reasoning, HRM is worth a look. It's a compact recurrent model that splits work between a slow high-level planner and a fast low-level executor, solving tasks like Sudoku, maze pathfinding, and ARC puzzles in a single forward pass without pretraining or explicit chain-of-thought supervision. The design directly tackles the brittleness and data hunger of standard CoT prompting, and the repo supports hands-on probing: you can train the small demos on a single laptop GPU and check the released checkpoints for harder benchmarks. This is a research codebase, not a production service, so treat it as a reference implementation and a good starting point for work on sample-efficient reasoning architectures.