Tiny Artificial-Intelligence Model Beats Giants in Logic Challenge

A miniature AI system known as the Tiny Recursive Model (TRM) has surprised researchers by outperforming some of the world’s most advanced language models in a demanding visual-reasoning test. The model succeeded on the ARC-AGI benchmark, a set of logic puzzles designed specifically to challenge machine intelligence.

Developed by Alexia Jolicoeur-Martineau of Samsung’s Advanced Institute of Technology in Montreal, the TRM uses only seven million parameters, making it nearly ten thousand times smaller than frontier large language models. Unlike traditional systems that rely on massive datasets, TRM trains on roughly a thousand examples for each puzzle type, learning tasks such as sudokus and mazes through simple trial-and-refinement steps.

Experts say the achievement shows the growing promise of compact, specialised models. François Chollet, creator of the ARC-AGI test, called the results significant and expects more research to build on this approach. The method differs sharply from large language models, which often struggle with unpredictable puzzles because they depend on statistical patterns learned from billions of documents.

The TRM borrows ideas from hierarchical reasoning, an iterative strategy where the model repeatedly improves its guesses. This allows it to solve new puzzles by looping through up to sixteen refinement steps, mimicking a small but focused reasoning process.

Researchers caution that the model is highly specialised and not a substitute for general purpose AI. Still, its success suggests that breakthroughs in reasoning may come not only from scaling up but from smarter, smaller designs. The model’s code is now publicly available, inviting wider experimentation and further advances in efficient, logic-driven AI.