Energy-Based Reasoners: Your Recurrent Reasoner Should Be an Energy-Based Model
The authors report 99.9% accuracy on Sudoku-Extreme using 345× less inference compute than the next-best model.
Energy-Based Reasoners learn a function that scores latent reasoning states, then use its gradient to guide repeated refinement and signal when to stop. Exploration-based training helps shape that function across a broader range of states. Across four algorithmic reasoning benchmarks, the models match or outperform prior recurrent reasoners, with comparable training compute. On every multi-solution benchmark tested, they achieve state-of-the-art coverage of distinct valid solutions while retaining near-100% single-sample accuracy.
HF Daily Papers' note
Energy-Based Reasoners learn a function that scores latent reasoning states, then use its gradient to guide repeated refinement and signal when to stop. Exploration-based training helps shape that function across a broader range of states. Across four algorithmic reasoning benchmarks, the models match or outperform prior recurrent reasoners, with comparable training compute. On every multi-solution benchmark tested, they achieve state-of-the-art coverage of distinct valid solutions while retaining near-100% single-sample accuracy.
HF Daily Papers' note
score 5