ResearchBenchmarksMath Question Answering on AIME 2025Follow51.7AccuracyICRL49.20449.85250.551.148Mar 9, 2026Evaluation ResultsMethodMethodLinksAccuracyICRLModel=Qwen3-8B, SFT=✗Model=Qwen3-8B, SFT=✗2026.0351.7ReToolModel=Qwen3-8B, SFT=✓Model=Qwen3-8B, SFT=✓2026.0349.3