Recursive Reasoning for Large Language Models: Scaling Compute at Inference

We introduce the Tiny Recursion Model (TRM), a lightweight architecture that decouples model size from reasoning depth by enabling recursive refinement of an internal latent state during inference. Instead of relying solely on parameter count, TRM dynamically increases computational depth by performing multiple recursive “thinking iterations” before emitting output tokens. This allows small models to demonstrate deep reasoning ability through inference-time compute scaling. We evaluate TRM on GSM8K, MATH, and algorithmic tasks such as Sudoku, showing significant gains over non-recursive baselines. This work highlights recursion as a promising alternative to scaling model size for efficient reasoning.