ResearchBenchmarksScience Question Answering on SciQ (val)Follow91.7AccuracyIA2 → SFT87.5488.6289.790.78Sep 26, 2025Evaluation ResultsMethodMethodLinksAccuracyExpected Calibration Error (ECE)IA2 → SFTModel=Qwen3-4B-Base, N...Model=Qwen3-4B-Base, N=4, Source Dataset=SciQr2025.0991.75IA2 onlyModel=Qwen3-4B-Base, N...Model=Qwen3-4B-Base, N=4, Source Dataset=SciQr2025.0990.49SFT onlyModel=Qwen3-4B-Base, N...Model=Qwen3-4B-Base, N=4, Source Dataset=SciQr2025.0988.37ICLModel=Qwen3-4B-Base, N...Model=Qwen3-4B-Base, N=4, Source Dataset=SciQr2025.0987.77