CRAM hits ~100% Sudoku-Extreme accuracy with just 30 training puzzles vs 700+ for looped models

IgorCarron · x · 2026-11-17

Ben Hoover reports that CRAM needs far less training data than looped models: 100% test accuracy on Sudoku-Extreme with only 30 puzzles, versus 700+ for larger looped models. The key insight: learning constraints is much easier than learning reasoning trajectories. Igor Carron frames it as a new kind of phase transition, recalling past work searching for phase transitions in Sudoku-solving algorithms.

Related event: IBM's CRAM Energy Model Trains 1,000x Faster on Sudoku(3 posts)→

Original post →

More from Research

Research channel →