ResearchBenchmarksReasoning on AIME25 and GPQA-DiamondReasoning on AIME25 and GPQA-DiamondFollow1.3978HV ValueSIP-BMM0.3276450.6054720.8832991.161126Dec 10, 2025Evaluation resultsMethodMethodLinksHV ValueCoveragePareto CountSIP-BMMGranularity=layer-wiseGranularity=layer-wise2025.121.3978—17DELLAGranularity=model-levelGranularity=model-level2025.121.3018—11TAGranularity=model-levelGranularity=model-level2025.121.2706—13DAREGranularity=model-levelGranularity=model-level2025.121.2572—11BREADCRUMBSGranularity=model-levelGranularity=model-level2025.120.5666—2TIESGranularity=model-levelGranularity=model-level2025.120.3688—5