TRM: Recursive vs. Autoregressive Reasoning
Samsung’s Tiny Recursive Model (TRM) is a 7-million-parameter model that argues against a core assumption of modern AI. Instead of scaling parameter counts up, it uses a small recursive network that “thinks” in a loop, repeatedly refining its own answer. On certain abstract-reasoning benchmarks it beats models thousands of times its size — which makes it less a product than a demonstration about where reasoning ability actually comes from.
Two ways to solve a problem
A standard LLM generates a solution token by token. TRM drafts a whole solution and then revises it, again and again, before committing. That is the entire architectural difference, and it reshapes everything else.
| Dimension | TRM (recursive) | Standard LLM (autoregressive) |
|---|---|---|
| Core logic | Iterative refinement (draft-then-revise) | Next-token prediction |
| Generation | Drafts a full solution, then recursively improves it | Emits output sequentially, one token at a time |
| Parameter efficiency | Extremely high (7M params) | Low (billions to trillions of params) |
| Where compute goes | Test-time (recursion) | Training-time (scale) |
| Strongest on | Structured, abstract reasoning puzzles (ARC, Sudoku) | General language, conversation, creativity |
TRM — the abstract-reasoning specialist
TRM is purpose-built for problems that reward repeated analysis and self-correction:
- Strong on ARC-AGI. It reaches roughly 45% on ARC-AGI-1, ahead of reported results for much larger models — around 37% for Gemini 2.5 Pro and about 16% for DeepSeek-R1 on this specific, difficult task.
- Structured puzzles. It handles tasks with clear rules and geometric or symbolic patterns, reporting 87.4% on Sudoku-Extreme and 85.3% on Maze-Hard.
- Where it fits. Primarily a research model for exploring efficient architectures, and a fit for specialised solvers in domains like logistics, formal verification, and scientific discovery where iterative reasoning is the point.
Autoregressive LLMs — general-purpose breadth
Standard LLMs like ChatGPT and Gemini are built for breadth and fluency: unmatched at generating human-like text and holding conversations, and backed by huge parameter counts that store broad world knowledge across nearly any topic. They are the default for chatbots, content creation, summarisation, translation, and any general-purpose language interface.
The lesson, and how the loop works
TRM is not a general-purpose AI and cannot converse; its significance is architectural.
- Use a standard LLM for virtually all production language tasks, from support bots to marketing copy.
- Study TRM’s architecture for how to build efficient, specialised solvers. It shows that for certain problem classes, spending compute on test-time reasoning — more “thinking” steps — can beat simply adding parameters.
The core mechanism is TRM’s recursive loop: it alternates a “think” step (updating a latent scratchpad) with an “act” step (refining the current solution). Because training backpropagates through the entire loop, the model learns to self-correct and generalise its reasoning process.
Constraints and access
TRM is not a language model in the usual sense — it is a specialised solver for structured reasoning and lacks an LLM’s breadth. Its recursive process trades speed for depth, so inference can be slower as it runs multiple refinement passes. It is a research project: Samsung researchers have published the code on GitHub for others to replicate and build on. It is not offered as a commercial API or product.
