Probabilistic Tiny Recursive Model Boosts Reasoning Without Retraining
A new AI framework called Probabilistic Tiny Recursive Model (PTRM) enhances reasoning accuracy by injecting Gaussian noise into recursive steps, enabling parallel exploration of solution basins. The method uses the model's existing Q head for selection, requiring no retraining or task-specific augmentations. PTRM achieves substantial gains on benchmarks including Sudoku, addressing the limitation of deterministic recursion in Tiny Recursive Models (TRM) which can converge on suboptimal solutions. The paper is available on arXiv under ID 2605.19943.
Key facts
- PTRM injects Gaussian noise at each deep recursion step.
- PTRM uses the model's existing Q head for early stopping and selection.
- PTRM requires no retraining or task-specific augmentations.
- PTRM achieves substantial accuracy gains across benchmarks including Sudoku.
- TRM solves complex reasoning tasks with a fraction of parameters of LLMs.
- TRM's deterministic recursion can lead to suboptimal solutions.
- PTRM is a task-agnostic framework for test-time compute scaling.
- The paper is available on arXiv:2605.19943.
Entities
Institutions
- arXiv