Verified project record

sapientinc/HRM

hrm is a recurrent reasoning architecture that achieves deep computational depth using minimal training data without pre-training or Chain-of-Thought techniques. It provides high-level planning and low-level computation modules for single-pass reasoning across tasks such as Sudoku, Maze, and ARC-AGI-2.

Project overview

By delivering strong reasoning performance from only 1000 training samples and eliminating the need for pre-training or CoT, hrm directly addresses key brittleness, data scarcity, and latency bottlenecks inherent in current LLM reasoning approaches.

Project type: Model Development · Infrastructure
Deployment: Refer to project documentation
License: Apache-2.0

Best for

AI engineers and researchers exploring recurrent reasoning architectures that avoid Chain-of-Thought dependencies.
Teams needing to train reasoning models with minimal data samples.

Key capabilities

A recurrent architecture featuring high-level planning and low-level computation modules designed for single-pass reasoning.
Train a master-level Sudoku AI capable of solving extremely difficult puzzles.
Evaluate trained models using exact accuracy and specific ARC evaluation scripts.
Provides trained checkpoints for ARC-AGI-2, Sudoku, and Maze.

Limitations and risks

Late-stage overfitting may cause numerical instability during training and Q-learning.
As an adoption check, small-sample learning exhibits accuracy variance of around ±2 points.

Getting started

Running the project is complex and requires coding knowledge. You must manually install CUDA 12.6, compatible PyTorch builds, and FlashAttention. While a minimum of an RTX 4070 laptop GPU is necessary, full-scale experiments require an 8-GPU hardware setup.
After installing dependencies via requirements.txt, build the Sudoku dataset, and then start training via the command-line interface.

Evidence and sources

README: Inspired by the hierarchical and multi-timescale processing in the human brain, we propose the Hierarchical Reasoning Model (HRM), a novel recurrent architecture that attains sign…
README: Current large language models (LLMs) primarily employ Chain-of-Thought (CoT) techniques, which suffer from brittle task decomposition, extensive data requirements, and high latenc…
README: Ensure PyTorch and CUDA are installed. The repo needs CUDA extensions to be built.
README: Then install FlashAttention. For Hopper GPUs, install FlashAttention 3
README: pip install -r requirements.txt

Reviewed at: 2026-07-18T00:41:33Z