The AI industry has spent years proving that larger models can do more. A paper from Alexia Jolicoeur-Martineau of Samsung SAIL Montréal points in the opposite direction: on several demanding reasoning tasks, a network with just seven million parameters can outperform models many thousands of times larger.
The Tiny Recursive Model, or TRM, is not trying to become a miniature general-purpose chatbot. It is designed to solve problems where precise, repeated correction matters more than broad linguistic knowledge. Sudoku, long mazes, and abstract visual puzzles punish one early mistake, making them an uncomfortable fit for language models that commit to an answer token by token.
TRM begins with a question, an initial proposed answer, and a hidden reasoning state. A small network refines that internal state, updates the answer, and repeats the cycle as many as 16 times. Instead of producing a long written chain of thought, the model learns to improve its own compact representation of the problem.
That design resembles returning to a difficult puzzle with a better draft rather than extending one irreversible answer. Compute is spent on revision inside a narrow network, not on carrying billions of parameters through every step. The contrast gives the work significance beyond its benchmark scores.
One Tiny Network Replaces a Complex Hierarchy
The research builds on the Hierarchical Reasoning Model, which used two networks operating at different frequencies. HRM produced promising results, but its design relied on biological analogies and mathematical assumptions about convergence that were not guaranteed to hold. TRM removes the hierarchy and back-propagates through the complete recursive process.
Simplification produced a striking gain. In an ablation study on Sudoku-Extreme, changing the training approach raised accuracy from 56.5 to 87.4 percent. A two-layer network also generalised better than a four-layer version. On small specialised datasets, extra capacity can memorise training patterns instead of learning the underlying operation; being tiny may be part of the model’s advantage, not merely a cost-saving compromise.
The Benchmark Results Challenge Scale
Sudoku-Extreme contains only 1,000 training examples, yet TRM reached 87.4 percent test accuracy compared with HRM’s 55 percent. On Maze-Hard, which requires paths through 30-by-30 grids, it scored 85.3 percent against HRM’s 74.5 percent. These are tasks where iterative correction has a natural role and one invalid step can destroy an otherwise plausible solution.
The most provocative result came from the Abstraction and Reasoning Corpus. TRM scored 44.6 percent on ARC-AGI-1 and 7.8 percent on the harder ARC-AGI-2 using seven million parameters. The source compares the latter with 4.9 percent for Gemini 2.5 Pro. Benchmark conditions and system design must be compared carefully, but the result is still a sharp reminder that parameter count is not a universal measure of reasoning ability.
Recursive Reasoning Opens a More Efficient Path
TRM also simplifies Adaptive Computation Time, the mechanism deciding when an answer has improved enough to move to another training sample. It removes a second costly forward pass without materially reducing generalisation. That makes both the architecture and its training procedure leaner.
The research does not establish that tiny recursive networks can replace large language models. TRM is specialised, while LLMs bring extensive knowledge, language ability, and flexibility across tasks. The exciting possibility is a different division of labour: a broad model could interpret a problem and call a compact recursive specialist when exact search or correction is required.
That would change the efficiency debate. Progress would no longer depend solely on building a larger general model and spending more compute at inference. Samsung’s result suggests architecture can sometimes beat scale—and that a system small enough to overlook may be exactly the one worth watching.
