Recursion Is The Next Scaling Law In AI

Recursion at inference time provides a viable alternative to scaling model size for enhancing AI reasoning. Unlike standard LLMs that rely on one-shot feed-forward processes, Hierarchical Reasoning Models (HRM) and Tiny Recursive Models (TRM) utilize iterative computation to tackle incompressible tasks like Sudoku and sorting. By employing a deep equilibrium approach and truncated backpropagation through time, these models leverage latent states as a memory cache, enabling complex reasoning without the prohibitive memory costs of massive transformer architectures. The TRM approach further optimizes this by collapsing hierarchical networks into a single shared-weight module, achieving superior performance on benchmarks like the ARC Prize with significantly fewer parameters. Integrating these compute-efficient recursive methods into general-purpose models offers a promising strategy to overcome the inherent reasoning limitations of current token-based architectures.