The tiny model created by a Samsung researcher out-reasons giants

Through a self-improvement loop of drafting, rethinking, and refining answers, Alexia Jolicoeur-Martineau of Samsung unveiled the Tiny Recursion Model, a 7M parameter AI that outperforms DeepSeek R1 and Gemini 2.5 Pro in difficult reasoning. The specifics: TRM outperformed models thousands of times larger, achieving 45% on the infamously challenging ARC-AGI-1 and 8% on ARC-AGI-2. TRM drafts solutions and refines them through up to 16 cycles of internal reasoning and revision rather than creating answers token by token. The model updates its answer draft after six cycles of reviewing and refining its reasoning in a separate scratchpad. The findings were encouraging for the highly specialized puzzle question types seen in ARC, but they may not be applicable to various areas of reasoning. Research like TRM (and Sapient's HRM) demonstrates that clever architectural adjustments can level the playing field for small, effective models in the competition for billions of dollars' worth of computation and vast scale in AI models. Although puzzles are the main focus here, the idea may alter how labs with little funding approach AI research.