LESS IS MORE: RECURSIVE REASONING WITH TINY NETWORKS

Summary from the TLDR AI newsletter

A single Samsung researcher developed Tiny Recursive Model, which uses a 2-layer network that recursively improves its answer up to 16 times. It outperforms DeepSeek R1 and o3-mini-high on ARC-AGI by 10% to 30% despite having 0.01% of the parameters. The approach simplifies the Hierarchical Reasoning Model by eliminating complex mathematical theorems and dual-network hierarchies, and suggests that smaller networks with deeper recursion generalize better than larger ones without recursion.