ResearchBenchmarksMultimodal Reasoning on OOD MultimodalFollow64Average Sample Size (N)MV4.855220.210135.56550.9199Jul 2, 2026Evaluation ResultsMethodMethodLinksAverage Sample Size (N)Speedup FactorAccuracyMVModel=Qwen2.5VL-2.5-7B...Model=Qwen2.5VL-2.5-7B-Instruct, Stopping criteria=N=642026.0764157.8ACModel=Qwen2.5VL-2.5-7B...Model=Qwen2.5VL-2.5-7B-Instruct, Stopping criteria=Adaptive2026.0715.244.257.9CASModel=Qwen2.5VL-2.5-7B...Model=Qwen2.5VL-2.5-7B-Instruct, Stopping criteria=Confidence-based Adaptive2026.077.138.9857.9