arXiv 2510.04871

Less is More: Recursive Reasoning with Tiny Networks

By Alexia Jolicoeur-Martineau

Published 2025-10-06

Mindmap

Browse the paper's core ideas, clusters, and relationships in a structured outline.

Hierarchical Reasoning Model (HRM) is a novel approach using two small neural networks recursing at different frequencies. This biologically inspired method beats Large Language models (LLMs) on hard puzzle tasks such as Sudoku, Maze, and ARC-AGI while trained with small models (27M parameters) on small data (around 1000 examples). HRM holds great promise for solving hard problems with small networks, but it is not…

View the original paper on arXiv