Less is More: How Tiny Networks Outperform Giant LLMs on Hard Puzzles

Author(s): Gowtham Boyina

Originally published on Towards AI.

A deep dive into Tiny Recursive Models (TRM) — achieving 45% accuracy on ARC-AGI-1 with 7M parameters, outperforming models 10,000x larger

Large Language Models have revolutionized AI, but they struggle on certain types of hard reasoning tasks. Consider ARC-AGI — a benchmark of geometric puzzles designed to be easy for humans but challenging for AI. Despite years of progress, even state-of-the-art LLMs with massive parameter counts struggle to make significant headway, particularly on the newer ARC-AGI-2 benchmark.

The article explores the limitations of Large Language Models (LLMs) in solving complex reasoning tasks and introduces Tiny Recursive Models (TRM), a new architecture that performs better with significantly fewer parameters. Developed by Samsung SAIL, TRM achieves remarkable accuracy on the ARC-AGI benchmarks through deep supervision and a unique iterative refinement process. The approach challenges the common belief that larger models are necessarily better, demonstrating that effective architectural design can yield superior results even with minimal parameters. By training models to self-correct, TRM provides insights for addressing specific problem classes more efficiently.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI