Energy-Based Reasoners: 1.8M-param model hits 99.9% on Sudoku-Extreme with 345x less inference compute

LucaAmb · x · 2026-11-19

Researchers from Duke, UIUC, and Harvard introduce Energy-Based Reasoners (EBRs), which learn an energy function over latent reasoning states and perform looped updates as gradient descent on it — the energy scores candidate states, gradients guide refinement, and small gradients give a natural stopping signal, no external verifier needed. With exploration-based training, EBRs reach 99.9% on Sudoku-Extreme using 345x less inference compute than the next-best model, use 2.6x fewer training FLOPs, are 7-43x more sample-efficient than looped baselines, and achieve near-100% accuracy on multi-solution benchmarks plus SOTA coverage on N-Queens and Graph Coloring.

Related event: Energy-Based Reasoners Cut Inference Compute 345x, Solve 99.9% of Hard Sudoku(3 posts)→

Original post →

More from Research

Research channel →