sapientinc/HRM

Implements a dual-module recurrent architecture with separate high-level planning and low-level execution streams, enabling complex reasoning in a single forward pass without intermediate supervision. Achieves strong performance on abstract reasoning benchmarks (ARC, Sudoku, maze solving) with minimal parameters (27M) and training data (1K samples), using FlashAttention for GPU-optimized inference. Integrates with PyTorch, Weights & Biases for experiment tracking, and provides pre-trained checkpoints on Hugging Face.

12,358 stars. No commits in the last 6 months.