A researcher at Samsung’s AI lab in Montreal has published a paper describing a 7-million-parameter neural network that outperforms several large language models on abstract reasoning benchmarks. The Tiny Recursive Model (TRM), detailed in a paper on arXiv, achieves 45% accuracy on ARC-AGI-1 and 8% on ARC-AGI-2, scores that surpass those of DeepSeek-R1, OpenAI’s o3-mini, and Google’s Gemini 2.5 Pro, according to the paper.

TRM works by recursively refining its answers. It maintains two short-term memories: one for the current solution and another as a scratchpad for intermediate steps. At each iteration, it reviews the problem, its current answer, and its notes, then updates its output. An adaptive halting mechanism stops the process when improvements plateau. The approach uses a single two-layer network, in contrast to the Hierarchical Reasoning Model (HRM) that inspired it.

On other benchmarks, TRM scored 87% on Sudoku-Extreme and 85% on Maze-Hard. In an X post, Jolicoeur-Martineau says the model trains for under $500 using four H-100 GPUs over about two days. While TRM excels at grid-based puzzles, it is not designed for open-ended tasks like conversation or image generation. The paper suggests that recursive reasoning may offer a path to strong performance on structured problems without massive parameter counts.