ResearchBenchmarksParagraph-level Discrete Reasoning on DROPFollow80.4AccuracyBase80.08880.16980.2580.331Sep 29, 2025Evaluation ResultsMethodMethodLinksAccuracyBasebackbone=DS-8Bbackbone=DS-8B2025.0980.4RealSafebackbone=DS-8B, alignm...backbone=DS-8B, alignment=SFT-based2025.0980.1IPObackbone=DS-8Bbackbone=DS-8B2025.0980.1