Less Is More: Recursive Reasoning with Tiny Networks
13 hours ago
- #tiny-networks
- #AI-efficiency
- #recursive-reasoning
- Tiny Recursion Model (TRM) achieves 45% on ARC-AGI-1 and 8% on ARC-AGI-2 with a 7M parameter network.
- TRM simplifies recursive reasoning, avoiding reliance on large models or biological arguments.
- The model recursively improves answers by updating latent states and answers efficiently.
- Setup requires Python 3.10, Cuda 12.6.0, and specific pip installations including torch and wandb.
- Training scripts provided for ARC-AGI-1, ARC-AGI-2, Sudoku-Extreme, and Maze-Hard datasets.
- Training times vary, with ARC-AGI taking ~3 days and Sudoku/Maze tasks completing in <36 hours.
- Citations include TRM and the Hierarchical Reasoning Model (HRM) as foundational work.