Energy-Based Reasoners: 385x leaner inference
Original: Energy-Based Reasoners: 385x more efficient inference
Why This Matters
Dramatic inference efficiency gains matter as reasoning models scale to harder real-world tasks.
Researchers from Duke, UIUC, and Harvard introduce Energy-Based Reasoners (EBRs), a 1.8M-parameter model achieving 99.9% accuracy on Sudoku-Extreme with 345x less inference compute than the next-best model.
A joint team from Duke University, UIUC, and Harvard has published Energy-Based Reasoners (EBRs), a new class of recurrent reasoning model that replaces ad-hoc loop termination with principled gradient descent on a learned energy function. The key insight: existing looped models iterate without any internal measure of progress, so they can't tell when to stop or how to improve. EBRs fix this by scoring latent reasoning states with an energy function—lower energy signals better compatibility with the input—and using the gradient of that energy both to refine the state and to decide when to halt. Stopping happens automatically when the gradient norm falls below a threshold, so easy problems finish quickly and hard ones keep iterating. At 1.8M parameters, EBRs hit 99.9% accuracy on Sudoku-Extreme—a benchmark where the next-best model (Looped Flows, ~5-7M params, 97.9%) requires 345x more inference compute. On multi-solution tasks like N-Queens and Graph Coloring, EBRs achieve state-of-the-art coverage of distinct valid solutions while holding near-100% single-sample accuracy. The team also introduced exploration-based training to discover additional solution modes and shape the energy surface more broadly. Test-time scaling works two ways: more descent steps refine each answer, and more independent trajectories give the energy scorer more candidates to pick from. EBRs match EqR's best accuracy with 7x fewer iterations and 43x fewer samples.