Energy-Based Reasoners Achieve 99.9% Sudoku Accuracy with 345x Less Inference Compute

·2026.11.18 18:29

Key point

The new Energy-Based Reasoners (EBRs) match or outperform prior recurrent models on algorithmic benchmarks while using 345x less inference compute than the next-best model.

1 / 4

Details

  • Energy-Based Reasoners (EBRs) formulate looped updates as gradient descent on a learned energy function, providing explicit signals for refinement, stopping, and solution selection.
  • On Sudoku-Extreme, EBRs achieve 99.9% accuracy with 5 samples, surpassing EqR which requires 128 samples to reach 99.8%.
  • EBRs use 345x less inference compute than the next-best model and 10.9x fewer FLOPs than a fixed-depth EBR by using adaptive stopping based on energy gradient thresholds.
  • The model uses only 1.8M parameters, compared to 5-10M for baselines like TRM, EqR, and GRAM, while maintaining comparable training compute.
  • EBRs achieve state-of-the-art coverage of distinct valid solutions on multi-solution benchmarks, reaching 98.1% coverage on 10x10 N-Queens with 20 samples.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.