A 7M parameter model beating DeepSeek-R1, Gemini 2.5 Pro, and o3-mini at reasoning? 🤯
That’s what
@jm_alexia just showed at Mila — and the secret isn’t size, it’s recursion.
Tiny Recursive Models (TRMs) rethink reasoning from scratch: instead of brute-forcing scale, they learn to reason by improving their own answers step by step. No biological metaphors, no fixed-point math — just clean, recursive logic.
The result?
A toy-sized network outperforming giants on ARC-AGI, Sudoku, and Maze-Hard. Sometimes, less really is more.
And it all started with one goal:
#SaveTheCurve 😂
I’m Louis-François, PhD dropout, now CTO & co-founder at Towards AI. Follow me for tomorrow’s no-BS AI roundup 🚀
#AIResearch #DeepLearning #MachineLearning #TinyModels #RecursiveReasoning #TRM