A new paper titled 'Less is More: Recursive Reasoning with Tiny Networks' introduces the Tiny Recursive Model (TRM), a 7-million-parameter architecture inspired by the Hierarchical Reasoning Model (HRM). TRM uses latent reasoning instead of chain-of-thought, maintaining two recursively refined latent features: Z for internal reasoning and Y for the current solution embedding. Unlike HRM's two-module design, TRM uses a single shared-weight module applied repeatedly. It improves on HRM's one-step gradient approximation by backpropagating through a full cycle, removing the fixed-point convergence assumption. TRM also features adaptive halting, allowing it to dynamically adjust compute based on task difficulty. Despite its tiny size, TRM outperforms HRM by over 30% on Sudoku and beats much larger models like Gemini 2.5 Pro and Claude 3.7 on ARC-AGI and maze-solving benchmarks. The 'less is more' framing may partly reflect overfitting constraints from limited training data rather than a universal architectural principle.
Sort: