How Can A Model 10,000× Smaller Outsmart ChatGPT?

The Tiny Recursion Model (TRM) challenges the assumption that bigger AI models are smarter. With only 5-7 million parameters, TRM uses a recursive loop architecture — maintaining three state vectors (question, hypothesis, latent reasoning) — to iteratively refine answers rather than making a single forward pass. On the Sudoku-Extreme benchmark, TRM achieved 87.4% accuracy while GPT o3-mini, Claude 3.7, and DeepSeek R1 scored 0%. On ARC-AGI-1, TRM's 7M parameter model hit 44.6% accuracy, outperforming DeepSeek R1 (671B params, 15.8%) and Claude 3.7 (28.6%). A counterintuitive finding: doubling TRM's depth from 2 to 4 layers reduced accuracy from 87.4% to 79.5%, suggesting extra capacity encourages memorization over deduction. TRM also uses Adaptive Computation Time to dynamically decide when to stop iterating, allocating compute only where needed. The core thesis: depth in time beats depth in space.