CRAM energy model trains 1,000x faster, solving 99.996% of Sudoku-Extreme in 42ms
udmrzn · x · 2026-11-16
Researchers (Hoover et al.) unveiled CRAM, a learned version of Hopfield & Tank's 1980s energy that challenges looped-model reasoning.
Key results from the arXiv paper:
- Solves 99.996% of Sudoku-Extreme after just 42ms of training on one GPU, and 100% of Maze-Unique in 132ms — at least 1,000x faster than looped models
- Matches baseline accuracy on Sudoku-Extreme with at least 282x fewer inference FLOPs
- No backpropagation through time; adaptive inference via energy minimization, with interpretable weights that learn puzzle constraints
The authors caution that trivially solvable benchmarks like Sudoku and Maze are poor tests of looped reasoning scalability.
More from Research
- Claude describes the moment of being corrected; test qualia via explanatory power — Sauers_ · 2026-11-16
- Sakana AI releases JFinBench: 321 Japanese finance exam questions expose open-model gaps — SakanaAILabs · 2026-11-16
- TetraCoder: adaptive tetrahedral encoding beats Instant-NGP PSNR with fewer parameters — ssh4net · 2026-11-16
- MMFlash boosts vision-language speculative decoding with up to 3.58x lossless speedup — _akhaliq · 2026-11-16
- When AI gets smart enough, physical experiment speed becomes science's only bottleneck — VraserX · 2026-11-16
- Meta's e-means beats k-means at 1/10 the epochs, up to 45x GPU speedups — udmrzn · 2026-11-16