Energy-Based Reasoners: 385x more efficient inference
The reported result is 99.9% accuracy on Sudoku-Extreme with 345× less inference compute—not the headline’s 385×.
Energy-Based Reasoners learn a score over latent reasoning states and use its gradient to refine answers and decide when to stop. Adaptive stopping preserves that accuracy while using 10.9× fewer FLOPs than a fixed-depth EBR. The 1.8-million-parameter model reaches 99.9% with five samples, compared with EqR’s 99.8% with 128. Across the multi-solution benchmarks, EBRs also find a greater share of distinct valid solutions than the reported baselines.
HN · Inference's note
Energy-Based Reasoners learn a score over latent reasoning states and use its gradient to refine answers and decide when to stop. Adaptive stopping preserves that accuracy while using 10.9× fewer FLOPs than a fixed-depth EBR. The 1.8-million-parameter model reaches 99.9% with five samples, compared with EqR’s 99.8% with 128. Across the multi-solution benchmarks, EBRs also find a greater share of distinct valid solutions than the reported baselines.
HN · Inference's note
score 5