CRAM turns Sudoku into energy descent: 99.996% of Sudoku-Extreme solved after 42 ms of training, 1,000 times faster than looped models
Synopsis
A trainable energy model called CRAM turns reasoning into repeated descent on an energy function whose weights are learned directly from solved puzzles. It solves 99.996% of Sudoku-Extreme test puzzles and every Maze-Unique test maze, with training on one GPU taking 42 ms and 132 ms respectively, at least 1,000 times faster than looped models. It also cautions that high scores on puzzles like Sudoku and mazes do not indicate how looped reasoning methods will scale to harder tasks.
Interpretation
Reasoning need not come from applying the same block of layers over and over; it can come from minimizing an energy function learned from data, and this energy model, CRAM, reaches higher accuracy with far shorter training on constraint puzzles. Earlier looped reasoning models repeatedly apply a layer or block of layers to produce an answer, work on Sudoku and mazes, but train slowly, can be unstable, and use many parameters and FLOPs. CRAM solves 99.996% of Sudoku-Extreme test puzzles against 99.87% for the most accurate looped baseline, TRMP, and 100% of Maze-Unique against 99.23% for TRMP. CRAM's fit times are 42 ms and 132 ms, while the looped baselines' fit times range from about 13 minutes to more than twenty hours.
Reasoning proceeds by repeatedly updating each cell's belief over candidate values along the energy gradient, until the energy falls below a threshold fitted from training solutions or no single-cell change lowers it. The energy is written as a mean-field free energy with pairwise costs between cells plus a per-cell bias, extending the continuous Hopfield network energy to multi-state neurons; training uses a pseudolikelihood objective computed directly at labeled solutions, without backpropagating through the reasoning trajectory. In every Sudoku-Extreme training run, CRAM learned to place positive equal-digit costs on exactly the 810 cell pairs sharing a row, column or box and zero elsewhere; on Maze-Unique, its 18 learned costs make each search move turn off one dead-end cell, which is dead-end filling.
The energy supplies a natural stopping test, so compute follows puzzle difficulty, and the search never stops at an incorrect board. Looped models usually add a separate halting network or stop on a gradient-norm rule, and these rules stop at an incorrect board on some puzzles. Binning 10,000 Sudoku-Extreme test puzzles by the dataset's difficulty rating, CRAM's median compute grows 22-fold from the easiest bin to the hardest; the halting rules of HRM, TRM, EqR, TRMP and EBR each stop at an incorrect board on 0.05% to 3.54% of the puzzles they stop on, versus 0 for CRAM.
Perspective
The demonstrated scope is four constraint puzzles, Sudoku-Extreme, Maze-Unique, N-Queens and graph coloring on 8 and 10 vertices, all of which CRAM's pairwise energy can represent exactly; all fit times are wall-clock measurements on one H200. On Maze-Hard, plain CRAM solves none of the test mazes, while a version with one added hidden layer solves 999 of 1,000 test mazes in each of three training runs.
High accuracy only shows that the model learned a few pairwise rules between answer cells, so high scores on these benchmarks do not indicate performance on harder tasks such as ARC-AGI. In the extremely data-limited regime CRAM converges to many incorrect solutions, while TRM solves the most test puzzles with the fewest training puzzles. Fit times were measured on a shared cluster and vary with concurrent work on a node.
