Generative Recursive Reasoning Models Show Superior Scaling on Sudoku

anshulkundaje · x · 2026-08-02

Researchers have found that combining exploration with latent reasoning models effectively encourages diversity in reasoning latents and thoughts.

This approach, known as Generative Recursive Reasoning (GRAM), demonstrates excellent test-time scaling on tasks like Sudoku by utilizing parallel trajectory sampling. The associated paper (arXiv:2605.19376) suggests this framework enables end-to-end training, which could significantly improve out-of-distribution (OOD) generalization.

Original post →

More from Research

Research channel →