List your company

Combinatorial Reasoning on Sudoku

100AccuracyOracle policy
88.216891.275994.33597.3941May 29, 2026

Evaluation results

MethodLinks
Oracle policyTraining protocol=Policy-only training2026.05100
PolicyTraining protocol=Joint training2026.0592.87
BaselineTraining protocol=Joint training2026.0592.72
High conf.Training protocol=Joint training2026.0592.68
MarginTraining protocol=Joint training2026.0591
PolicyTraining protocol=Policy-only training2026.0590.82
High conf.Training protocol=Policy-only training2026.0589.84
MarginTraining protocol=Policy-only training2026.0588.67