ResearchBenchmarksMathematical Reasoning on AIME25 (Average Accuracy (%))Follow19.1Average AccuracyGRPO7.45210.47613.516.524Dec 1, 2025Evaluation ResultsMethodMethodLinksAverage AccuracyGRPOBackbone=Qwen3-4B-Base...Backbone=Qwen3-4B-Base, Sample Count=Avg@16, Evaluation Domain=In-domain2025.1219.1ReProBackbone=Qwen3-4B-Base...Backbone=Qwen3-4B-Base, Sample Count=Avg@16, Evaluation Domain=In-domain, Base Method=GRPO2025.1216.7OriginalBackbone=Qwen3-4B-Base...Backbone=Qwen3-4B-Base, Sample Count=Avg@16, Evaluation Domain=In-domain2025.127.9