Tiny Recursive Model (TRM) achieves higher ARC-AGI accuracy than LLMs with 7M parameters

Researchers propose the Tiny Recursive Model (TRM), a simplified recursive reasoning approach using a single two-layer network, which outperforms larger models on complex puzzle tasks. TRM utilizes only 7 million parameters and is trained on minimal data, demonstrating significantly higher generalization than the previously proposed Hierarchical Reasoning Model (HRM).

• TRM achieves 45% test accuracy on ARC-AGI-1 and 8% on ARC-AGI-2.
• These results exceed the performance of large language models such as Deepseek R1, o3-mini, and Gemini 2.5 Pro.
• The model uses less than 0.01% of the parameters required by those competing LLMs.

This work suggests that hard problems can be solved effectively with very small networks, offering a more efficient alternative to large-scale language models for specific reasoning tasks.