Less is More: Recursive Reasoning with Tiny Networks

The paper operates within the field of AI reasoning models. It challenges the dominant paradigm of Large Language Models (LLMs), which leverage massive scale and data to solve problems. The paper identifies a crisis in this paradigm, noting LLMs' struggles with specific hard puzzle tasks. It proposes a revolutionary alternative paradigm centered on recursive reasoning with small, efficient networks. This work refines a nascent revolutionary model (HRM) into a much simpler and more powerful version (TRM), presenting it as a superior solution for the anomalies that plague the LLM paradigm.

This preprint is classified as Model Revolution because it directly confronts anomalies in the prevailing LLM paradigm—namely, its failure on hard reasoning tasks. The authors do not merely offer an incremental improvement but propose a fundamentally different and simpler architecture, the Tiny Recursive Model (TRM). By demonstrating that a tiny, recursive network can significantly outperform massive LLMs on these specific problems, the paper offers a compelling new framework. This act of proposing a more elegant and powerful solution to a known crisis is the hallmark of a scientific revolution in progress.