Hopfield-Era Energy Model CRAM Solves Reasoning Puzzles 1,000x Faster Than Looped Transformers

The Three-Question Framework

What Happened?

In arXiv preprint 2611.17309 posted November 12, 2026, Prefrontal researchers introduced CRAM—a learned 1980s Hopfield-Tank pairwise energy model that hits 99.996% to 100% accuracy on Sudoku-Extreme after 42 ms of H200 training.

Why It Matters

Training small explicit energy weights via local objectives eliminates slow backpropagation through time (BPTT), running at least 1,000x faster to train and using 282x fewer inference FLOPs while exposing puzzle benchmarks as saturated.

What Could Happen Next

The authors caution against relying on Sudoku or simple mazes to judge looped reasoning scale-up, noting latent extensions like those tested on Maze-Hard are needed before tackling complex benchmarks such as ARC-AGI.

Executive Key Takeaways

On one Nvidia H200 GPU across three seeds, CRAM solved 100.00% of Sudoku-Extreme with 7.7K nonzero parameters versus 5M to 27M parameters and 1.2 to 9.0 hours of training in looped baselines.

On Maze-Unique, CRAM compressed down to 18 nonzero parameters, trained in 132 milliseconds, and used 8.37e5 FLOPs per puzzle against baselines spending 2.0 to 23.2 hours.

CRAM replaces separate learned halting heads with native energy-threshold stopping and parallel noisy Gibbs chains (`NChains=4` on Sudoku) that adapt compute to difficulty.

Because learned pairwise weights match human-interpretable rule constraints—with a hand-set energy version also hitting 100%—the authors argue puzzle benchmarks flatter looped architectures.

Source: arXiv:2611.17309v1 ('Classical Energies Learn to Reason 1,000x Faster than Looped Models', submitted November 12, 2026 by Benjamin Hoover, Hugh Blayney, and James C. R. Whittington at Prefrontal in London).

Looped neural networks repeatedly pass states through shared transformer blocks to simulate thinking time, adding external halting modules and differentiating through entire rollout trajectories via unstable truncated BPTT. Prefrontal's team flipped that paradigm by resurrecting Hopfield and Tank's 1980s constraint-energy formulation: CRAM defines a mean-field free energy pairing local cell entropy with learned pairwise compatibility matrices and clamps known puzzle clues before running focused Gibbs updates.

Benchmark comparisons underscored extreme efficiency gaps on a single H200 GPU. Where prior looped reasoners spanned 54.62% on HRM (27.0M parameters, 2.20e11 FLOPs, 1.2 hours), 85.73% on TRM (5.0M parameters, 2.5 hours), and 99.85% on EBR (1.8M parameters, 3.37e11 FLOPs, 2.8 hours), CRAM converged via pseudolikelihood without differentiating through any solver in 0.04 seconds at 2.02e9 FLOPs per puzzle.

Equally revealing was CRAM's inspectability: its trained interaction matrices visibly encode exact row, column, and box exclusion rules and learn clean Sudoku weights from an order of magnitude fewer examples than looped models. Using hand-set weights directly translated from puzzle rules also scored 100.00% on both Sudoku-Extreme and Maze-Unique.

The researchers stressed an important boundary on what these numbers imply for general AI reasoning: CRAM's pure pairwise graph cannot represent higher-order abstractions needed for open-ended benchmarks like ARC-AGI on its own, though a small hidden-layer latent variant handled harder mazes in Appendix F.

The Three-Question Framework

What Happened?

Why It Matters

What Could Happen Next

Executive Key Takeaways

Primary Sources & Citations