Energy-Based Reasoners achieve 385x more efficient inference
github.io

Energy-Based Reasoners achieve 385x more efficient inference

The gist

A Duke, UIUC and Harvard team released Energy-Based Reasoners, which solve 99.9% of Sudoku-Extreme puzzles with 345× less inference compute.
The approach swaps fixed looping for energy-guided stopping, a shift for efficient AI reasoning.

How it unfolded

  1. Nov 12, 2026A preprint, "Classical Energies Learn to Reason 1,000× Faster than Looped Models," presents CRAM, a trainable version of the Hopfield & Tank energy, solving 99.996% of Sudoku-Extreme and 100% of Maze-Unique test puzzles after 42 ms and 132 ms of training on one GPU.
  2. Nov 17, 2026EBR co-author William Chen posts on X announcing Energy-Based Reasoners, citing 345x less inference compute and 2.6x fewer training FLOPs than the best baselines.
  3. Nov 17, 2026Co-author Yilun Du posts on X that EBRs reason by taking gradient steps on a learned energy function and stop when the gradient gets small, solving 99.9% of Sudoku-Extreme puzzles with 1.8M parameters.

Key terms

inference compute
The process by which an AI system applies rules to known facts to derive new conclusions.
gradient descent
A first-order iterative optimization method that minimizes a differentiable function by repeatedly stepping in the direction of the negative gradient.
BPTT
A gradient-based technique for training certain recurrent neural networks by unrolling them over time.

Source: Wikipedia

Sources

Read the full story in the app Gistline — free on the App Store