ResearchBenchmarksGameCWM Generation on games (held-out)Follow66.7Mean Verification ScoreGPT-4o46.52451.7625762.238May 23, 2026Evaluation ResultsMethodMethodLinksMean Verification Score90% CI (Lower Bound)95% CI (Lower Bound)GPT-4oN=175N=1752026.0566.764.564.1SFT+GRPON=175N=1752026.0553.251.351GRPON=175N=1752026.0550.448.648.3SFTN=175N=1752026.054946.846.4BaseN=175N=1752026.0547.344.844.3