ResearchNov 18, 2026, 4:29 AM EST

Energy-Based Reasoners hit 99.9% on Sudoku with 345x less inference compute

Researchers at Duke, UIUC and Harvard introduced Energy-Based Reasoners, a class of reasoning models that learn an energy function over latent reasoning states and treat looped updates as gradient descent on that energy. The energy score lets the model choose among candidate answers without an external verifier, while its gradient indicates how to refine a state and when to stop looping. With 1.8M parameters, the method reached 99.9% accuracy on Sudoku-Extreme using 345× less inference compute than the next-best model, and state-of-the-art coverage on multi-solution N-Queens and Graph Coloring benchmarks. The work is presented in a paper and project page; no code release or deployment is mentioned.

Why it matters

Reasoning models that carry their own scoring and stopping signal could cut the compute spent on long chains of refinement, since easy problems can halt early instead of running a fixed number of steps.

Sources

Written by AIToolMaps from the source above and checked against it sentence by sentence.

Comments

No comments yet