ResearchBenchmarksMulti-token generation on HMathAFollow60.4AccuracyICL19.42430.06240.751.338Sep 26, 2025Evaluation ResultsMethodMethodLinksAccuracyICLBackbone=Qwen3-4B-Base...Backbone=Qwen3-4B-Base, N=42025.0960.4IA2 → SFTBackbone=Qwen3-4B-Base...Backbone=Qwen3-4B-Base, N=42025.0955.3SFT onlyBackbone=Qwen3-4B-Base...Backbone=Qwen3-4B-Base, N=42025.0950.4IA2 onlyBackbone=Qwen3-4B-Base...Backbone=Qwen3-4B-Base, N=42025.0947.8w/o ICLBackbone=Qwen3-4B-Base...Backbone=Qwen3-4B-Base, N=42025.0921