Beyond Bigger Models: Recursion As The Next Scaling Law In AI

Recursion offers a path to improve AI reasoning performance by enabling iterative computation at inference time rather than relying solely on increasing model size. Traditional LLMs operate as one-shot feed-forward processes, which limits their ability to solve incompressible problems like sorting or complex puzzles. Hierarchical Reasoning Models (HRM) and Tiny Recursive Models (TRM) address this by utilizing latent memory states—or "carries"—that allow the model to refine its output through repeated recursive steps. These architectures demonstrate that truncated backpropagation through time is sufficient for training, significantly reducing parameter requirements while achieving state-of-the-art results on challenging tasks like the ArcPrize. By integrating these efficient recursive methods with the robust semantic representations found in large-scale models, AI systems can achieve superior reasoning capabilities without the computational overhead of massive parameter scaling.