ResearchBenchmarksMulti-step Reasoning on CheckmateFollow90AccuracyeMoT47.56858.58469.680.616Jun 1, 2026Evaluation ResultsMethodMethodLinksAccuracyeMoTModels=Qwen-32BModels=Qwen-32B2026.0690BoTModels=GPT-4Models=GPT-42026.0686.4ToTModels=GPT-4Models=GPT-42026.0649.2