Guided Stochastic Exploration Boosts Recursive Model Accuracy on Sudoku-Extreme
A new study on arXiv (2605.25230) presents a method called guided stochastic exploration that aims to improve how recursive neural networks perform inference. The authors argue that the deterministic nature of these models essentially acts as a noise-free estimation of inference across hidden reasoning paths. By adding random variations and using the model's early-stopping feature to tweak candidate paths in real-time, they offer diagnostics that don't need labels, such as local stability, guide alignment, and cloud-token entropy. These indicators help predict the method's success and pinpoint reliable results. Impressively, this technique increases the exact-solve accuracy on the Sudoku-Extreme benchmark from 85.9% to 98.0% without any additional training.
Key facts
- Paper arXiv:2605.25230 proposes guided stochastic exploration for recursive neural architectures.
- Deterministic recursion is interpreted as the one-particle, zero-noise limit of approximate inference.
- Stochastic perturbations propose neighboring reasoning trajectories.
- The model's early-stopping head reweights trajectories online.
- Three label-free diagnostics are introduced: local stability, guide alignment, and cloud-token entropy.
- Sudoku-Extreme exact-solve accuracy improves from 85.9% to 98.0%.
- Method requires no additional training.
Entities
Institutions
- arXiv