ResearchBenchmarksMaximum-likelihood infilling on Cosmopedia 30 promptsFollow184Mean RewardGRPO182.8248183.1299183.435183.7401May 26, 2026Evaluation ResultsMethodMethodLinksMean RewardBest-of-8 ScoreVarianceEntropyPrefix Cross-Entropy (CE)Text Cross-Entropy (CE)GRPOTraining step=8kTraining step=8k2026.05184181.5313.550.1219.89164.11FrostTraining step=8kTraining step=8k2026.05182.87172.7781.870.5821.18161.69