Classical Energies Learn to Reason 1,000x Faster than Looped Models
CRAM solved nearly every test puzzle after milliseconds of training on one GPU.
The authors report 99.996% accuracy on Sudoku-Extreme after 42 milliseconds of training and 100% on Maze-Unique after 132 milliseconds. CRAM learns a version of Hopfield and Tank’s 1980s energy function, solving puzzles by descending that energy without differentiating through a solver during training. It also matched the strongest reported baseline looped-model accuracy on Sudoku-Extreme using at least 282 times fewer inference FLOPs. The authors caution that these easily solvable benchmarks offer limited evidence for the scalability of looped reasoning.
ArXiv · AI/CL/LG's note
The authors report 99.996% accuracy on Sudoku-Extreme after 42 milliseconds of training and 100% on Maze-Unique after 132 milliseconds. CRAM learns a version of Hopfield and Tank’s 1980s energy function, solving puzzles by descending that energy without differentiating through a solver during training. It also matched the strongest reported baseline looped-model accuracy on Sudoku-Extreme using at least 282 times fewer inference FLOPs. The authors caution that these easily solvable benchmarks offer limited evidence for the scalability of looped reasoning.
ArXiv · AI/CL/LG's note
score 5