ResearchBenchmarksMathematical Reasoning on AIME25Follow15.6AccuracyConSPO5.3047.97710.6513.323May 13, 2026Evaluation ResultsMethodMethodLinksAccuracyConSPOBackbone=Qwen3-4B-BaseBackbone=Qwen3-4B-Base2026.0515.6DisCOBackbone=Qwen3-4B-BaseBackbone=Qwen3-4B-Base2026.0512.2Dr. GRPOBackbone=Qwen3-4B-BaseBackbone=Qwen3-4B-Base2026.0511.4GRPOBackbone=Qwen3-4B-BaseBackbone=Qwen3-4B-Base2026.0510.7DAPOBackbone=Qwen3-4B-BaseBackbone=Qwen3-4B-Base2026.0510.1Base ModelBackbone=Qwen3-4B-BaseBackbone=Qwen3-4B-Base2026.055.7