Energy-Based Reasoners (EBRs) add an explicit progress measure to recurrent models that refine a latent state over repeated steps. A learned energy function scores how compatible each state is with the input, and gradient descent on that energy guides updates. The energy can rank candidate answers, while the size of its gradient indicates when refinement has settled and a trajectory can stop. The researchers also use exploration-based training to shape the energy landscape across a wider range of states and help the model discover distinct valid solutions.
On algorithmic reasoning tasks, a 1.8-million-parameter EBR reaches 99.4% single-sample accuracy on Sudoku-Extreme and 87.4% on Maze-Hard. With five trajectories, it achieves 99.9% and 88.3%, respectively, outperforming larger models that may use many more samples. Adaptive stopping preserves Sudoku accuracy while reducing compute 10.9-fold versus fixed-depth EBR inference; the reported comparison with EqR reaches the same best accuracy with 345 times fewer FLOPs. EBRs also produce strong results on tasks with multiple solutions: across 20 samples, they find 92.0% of distinct solutions for 10×10 N-Queens and 54.5% for 10-vertex graph coloring, while maintaining near-perfect single-sample validity. These results support energy-guided refinement as a way to combine accuracy, efficient stopping, and solution diversity.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.