Classical Energies Reason 1,000x Faster Than Looped Models

Author

AI News Editorial

Published

2026-11-15 10:15

Benjamin Hoover, Hugh Blayney and James C. R. Whittington ask a deliberately old question in a deliberately new paper: what if the energy function that powered neural networks in 1985 could be learned? Their answer, CRAM, is a trainable version of Hopfield & Tank’s energy, submitted on 12 November (arXiv 2611.17309, cs.AI/cs.LG).

The numbers are hard to ignore. After 42 milliseconds of training on a single GPU (132 ms for the second task), CRAM solves 99.996% of Sudoku-Extreme and 100% of Maze-Unique test puzzles. The authors put the comparison bluntly: at least 1,000x faster than looped models on these tasks, and at the highest looped-model accuracy on Sudoku-Extreme, CRAM uses at least 282x fewer inference FLOPs.

Why energy descent beats looping. Looped models re-run a block of layers until an answer settles — expensive per token, unstable in training, and heavy on parameters and FLOPs. Descent on an energy sidesteps most of that: it comes with built-in early stopping, compute that scales with problem difficulty rather than a fixed loop count, training that never differentiates through a solver, and weights that stay interpretable as the learned constraints of the puzzle rather than as opaque transformer circuitry.

The paper’s sting is aimed at benchmarks, not just models. Sudoku-Extreme and Maze-Unique are the current favourites of the looped-reasoning literature, and CRAM’s authors use that against the field: if a 1980s energy function cracks them in under a tenth of a second, then puzzles of this difficulty are measuring engineering effort, not reasoning capability. Their explicit recommendation is a caution against using trivially solvable puzzles to test whether looped reasoning techniques scale.

Read it next to the loop economics. Separate work on weight looping concluded that post-training loops only pay for small models — for larger ones the (N+m)/N per-token surcharge buys nothing. CRAM points the same way from the other end: if you want iterative compute, an energy you descend is cheaper, more stable and more legible than a transformer you re-apply.

The open question is transfer. Sudoku and Maze are structured, constraint-satisfiable, and closed-world; an energy that learns constraints beautifully may not survive open-ended text. The paper’s most valuable contribution for now is the second conclusion: a benchmark that a classical method solves in 42 milliseconds was never measuring what we thought it measured.

Sources: arXiv:2611.17309 · arXiv listing: cs.LG recent