Скачать или смотреть Outstanding Paper Review: Less is More - Tiny Recursive Models

The video introduces the Tiny Recursion Model (TRM), a new approach to recursive reasoning that uses a single, small neural network to tackle complex puzzle tasks. TRM is proposed as a simpler, more parameter-efficient, and superior alternative to the existing Hierarchical Reasoning Model (HRM), which uses two small networks and relies on complex theoretical justifications and biological arguments. TRM achieves significantly higher generalization on hard problems like Sudoku and ARC-AGI while using less than 0.01% of the parameters compared to large language models (LLMs) and outperforming HRM across several benchmarks. The paper details how TRM simplifies the recursive process by eliminating the need for complex elements like the Implicit Function Theorem and multiple networks, instead focusing on deep supervision and efficient single-network recursion. Ablation studies support TRM's design choices, showing that less is more through the optimal use of a single tiny network and a streamlined training process.