Hacker News AIT2·78 pts
- Casts looped reasoning updates as gradient descent on a learned energy function, with energy value serving as the stopping criterion
- Reaches 99.9% accuracy on Sudoku-Extreme with 345x less inference compute than the next-best model
- Introduces exploration-based training to shape the energy landscape over a broader set of states
- Achieves state-of-the-art coverage of distinct valid solutions on multi-solution benchmarks with near-100% single-sample accuracy