ResearchBenchmarksReasoning trace quality evaluation on CosmosQAFollow2.1Grammar ScoreCRAFT1.581.7151.851.985Apr 15, 2026Evaluation ResultsMethodMethodLinksGrammar ScoreRepetition Penalty (Step)Repetition Penalty (Word)CRAFTModel=o4-miniModel=o4-mini2026.042.122.8CRAFTModel=GPT-5.4-nanoModel=GPT-5.4-nano2026.041.61.71.7