ResearchBenchmarksMulti-token generation on SciQFollow40.8AccuracyIA2 → SFT5.54414.69723.8533.003Sep 26, 2025Evaluation ResultsMethodMethodLinksAccuracyIA2 → SFTBackbone=Qwen3-4B-Base...Backbone=Qwen3-4B-Base, N=42025.0940.8ICLBackbone=Qwen3-4B-Base...Backbone=Qwen3-4B-Base, N=42025.0937.5SFT onlyBackbone=Qwen3-4B-Base...Backbone=Qwen3-4B-Base, N=42025.0935w/o ICLBackbone=Qwen3-4B-Base...Backbone=Qwen3-4B-Base, N=42025.0915.4IA2 onlyBackbone=Qwen3-4B-Base...Backbone=Qwen3-4B-Base, N=42025.096.9