Research read

Energy-based reasoners hit 99.9% on Sudoku-Extreme with 345x less compute

Researchers from Duke, UIUC and Harvard propose Energy-Based Reasoners, looped reasoning models that learn an energy function to decide how to refine and when to stop. They report state-of-the-art results on algorithmic puzzles with a 1.8M-parameter model.

AE
Aidose Editorial
Nov 19, 2026

A new paper, "Energy-Based Reasoners: Your Recurrent Reasoner Should Be an Energy-Based Model," proposes recasting looped, or recurrent, reasoning models as energy-based models. The authors are William Chen and Zeyu Li of Duke University, Alexi Gladstone and Heng Ji of UIUC, and Yilun Du of Harvard University. The paper was posted to arXiv (2611.08977) on November 6, 2026 and drew wide attention this week.

The authors argue that existing looped reasoners lack an explicit measure of progress, which makes it hard to know when to stop iterating. Energy-Based Reasoners (EBRs) learn an energy function over latent reasoning states. Each loop step is gradient descent on that energy: the energy scores candidate states, its gradient guides refinement, and a small gradient serves as a stopping signal. The paper adds an "exploration-based training" method meant to discover more solution modes and shape the energy landscape.

With a 1.8-million-parameter model, EBRs reach 99.9% accuracy on Sudoku-Extreme with five samples and 99.4% with a single sample. The abstract says this uses 345 times less inference compute than the next-best model. Some online coverage gave the figure as 385x, but the paper says 345x. On Maze-Hard, EBRs score 88.3% with five samples and 87.4% with one.

On multi-solution tasks measured by coverage of distinct valid solutions (Cov@20), EBRs score 98.1% on 8x8 N-Queens versus 89.7% for the GRAM baseline, 92.0% on 10x10 N-Queens versus 57.5%, 88.6% on 8-vertex graph coloring versus 85.8%, and 54.5% on 10-vertex graph coloring versus 51.3%.

The results are on small algorithmic benchmarks, not language tasks. The idea most relevant to people building with AI is adaptive stopping: a learned signal for how much compute a problem needs, applied to test-time reasoning. Code is available at github.com/wwillchen/ebr.

AE
Aidose Editorial
Nov 19, 2026 · read
Back to News