Energy-Based Reasoners: 385x more efficient inference
Summary
This article introduces Energy-Based Reasoners (EBRs), a family of reasoning models that learn an energy function over latent states and perform looped updates via gradient descent. It reports dramatic efficiency gains (up to 345x fewer inference FLOPs) and strong accuracy on Sudoku-Extreme, N-Queens, and Graph Coloring, with adaptive stopping and multi-solution coverage. The work positions energy-based models as a promising direction for efficient, reliable reasoning in AI.