# Less is More: Recursive Reasoning with Tiny Networks

**Authors:** Alexia Jolicoeur-Martineau  
**arXiv:** [2510.04871](https://arxiv.org/abs/2510.04871)  
**Date:** 2025-10-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2510.04871v1/

## Key points

- The Hierarchical Reasoning Model (HRM) is an innovative approach in artificial intelligence research that utilizes two small neural networks operating at different frequencies and draws inspiration from biological systems.
- HRM has demonstrated superior performance over Large Language models (LLMs) in solving challenging puzzle tasks like Sudoku, Maze, and ARC-AGI using relatively modest resources - small models with 27 million parameters trained on a limited dataset of around 1000 examples.
- Despite its potential for solving complex problems with compact networks, the inner workings of HRM remain somewhat opaque and may not be fully optimized.
- The Tiny Recursive Model (TRM) has been proposed as a new approach to address these limitations, offering a simpler recursive reasoning strategy that surpasses HRM in terms of generalization capabilities.
- TRM achieves heightened performance using just a single tiny network consisting of only two layers and 7 million parameters, showcasing impressive test accuracy rates on benchmarks like ARC-AGI-1 and ARC-AGI-2 while outperforming most LLMs despite having significantly fewer parameters.
- The study by Alexia Jolicoeur-Martineau highlights the transformative potential of recursive reasoning models like TRM in advancing AI capabilities through minimalist approaches.

## Summary

In the realm of artificial intelligence research, a groundbreaking approach known as the Hierarchical Reasoning Model (HRM) has emerged. This innovative method utilizes two small neural networks operating at different frequencies and draws inspiration from biological systems. It has demonstrated superior performance over Large Language models (LLMs) in tackling challenging puzzle tasks like Sudoku, Maze, and ARC-AGI. Remarkably, HRM achieves this feat using relatively modest resources - specifically, small models with 27 million parameters trained on a limited dataset comprising around 1000 examples. While HRM shows immense potential for solving complex problems with compact networks, its inner workings remain somewhat opaque and may not be fully optimized. In response to this limitation, a new approach called the Tiny Recursive Model (TRM) has been proposed. TRM offers a simpler recursive reasoning strategy that surpasses HRM in terms of generalization capabilities. Surprisingly, TRM achieves this heightened performance using just a single tiny network consisting of only two layers and a mere 7 million parameters. The efficacy of TRM is underscored by its impressive test accuracy rates on challenging benchmarks such as ARC-AGI-1 and ARC-AGI-2. It outperforms most LLMs like Deepseek R1, o3-mini, and Gemini 2.5 Pro despite having less than 0.01% of their parameter count. This remarkable achievement highlights the power of minimalist approaches in artificial intelligence research and opens up new possibilities for efficient problem-solving with compact neural networks. The study conducted by Alexia Jolicoeur-Martineau sheds light on the transformative potential of recursive reasoning models like TRM in pushing the boundaries of AI capabilities.