Less Is More: Recursive Reasoning with Tiny Networks (arxiv.org)

Researchers propose a dramatic rethink of reasoning architectures with Tiny Recursive Model (TRM), a minimalist recursive network that outperforms much larger systems on hard puzzle tasks. Building on the Hierarchical Reasoning Model (HRM) — which uses two small networks recursing at different frequencies and required ~27M parameters — TRM pares the idea down to a single two-layer network with just 7M parameters. Trained on small datasets (~1k examples), TRM achieves 45% test accuracy on ARC-AGI-1 and 8% on ARC-AGI-2, and beats many contemporary LLMs (e.g., Deepseek R1, o3-mini, Gemini 2.5 Pro) while using under 0.01% of their parameter counts. Tasks include Sudoku, Maze, and ARC-style general intelligence puzzles. Why this matters: the results emphasize that inductive biases and algorithmic recursive structure can yield stronger generalization than brute-force scale, especially on combinatorial, algorithmic problems. TRM’s simplicity — a tiny 2‑layer network applying recursion to iteratively refine solutions — suggests a path toward highly data- and compute-efficient reasoning systems suitable for edge or embedded use. The approach is promising but not yet fully understood or necessarily optimal, highlighting opportunities to analyze how recursion and architectural priors enable algorithmic generalization and to extend these ideas to broader reasoning benchmarks.