'Tiny' AI model beats LLMs at logic test (www.nature.com)

Researchers have unveiled the Tiny Recursive Model (TRM), a 7‑million‑parameter neural model that outperformed some top large language models (LLMs) on the Abstract and Reasoning Corpus for Artificial General Intelligence (ARC‑AGI) — a suite of visual logic puzzles (sudokus, mazes, etc.) designed to stump machines. Unlike LLMs, TRM is highly specialized and not a language generator: it is trained from scratch on roughly 1,000 examples per puzzle type and solves unseen puzzles by iteratively guessing, comparing to the correct answer, and refining its solution up to 16 times. The model’s code is open source. The significance is twofold: TRM shows that compact, brain‑inspired architectures can learn robust reasoning strategies with orders of magnitude fewer parameters and data than frontier LLMs, challenging the notion that scale alone is the path to better reasoning. It was inspired by hierarchical iterative reasoning work and uses a small, recursive refinement loop rather than next‑token prediction over massive corpora. Caveats remain — TRM is narrowly scoped, must be retrained per problem class, and experts warn techniques that work at tiny scales may not translate directly to large models. Still, the work points to promising orthogonal approaches that could be integrated into or augment LLMs to improve systematic reasoning without trillion‑parameter costs.