A tiny recursive reasoning model achieves 45% on ARC-AGI-1 and 8% on ARC-AGI-2
7 hours ago
- #AI
- #Machine Learning
- #Recursive Reasoning
- Proposes Tiny Recursion Model (TRM), a recursive reasoning model with 7M parameters achieving 45% on ARC-AGI-1 and 8% on ARC-AGI-2.
- Challenges the notion that massive foundational models are necessary for solving hard tasks.
- Highlights the efficiency of recursive reasoning, showing 'less is more' in model size.
- TRM simplifies recursive reasoning, removing reliance on biological arguments or fixed-point theorems.
- Describes the recursive process: updates latent state and answer progressively to improve accuracy efficiently.