ResearchBenchmarksVision-language math reasoning on MathVerseFollow56.7Avg@32 AccuracyMHPO49.21251.15653.155.044Mar 14, 2026Evaluation ResultsMethodMethodLinksAvg@32 AccuracyMHPOBackbone=Qwen2.5-VL-7B...Backbone=Qwen2.5-VL-7B-Instruct2026.0356.7DAPOBackbone=Qwen2.5-VL-7B...Backbone=Qwen2.5-VL-7B-Instruct2026.0356.5SAPOBackbone=Qwen2.5-VL-7B...Backbone=Qwen2.5-VL-7B-Instruct2026.0356.3GRPOBackbone=Qwen2.5-VL-7B...Backbone=Qwen2.5-VL-7B-Instruct2026.0352.7GSPOBackbone=Qwen2.5-VL-7B...Backbone=Qwen2.5-VL-7B-Instruct2026.0352.1BaselineBackbone=Qwen2.5-VL-7B...Backbone=Qwen2.5-VL-7B-Instruct2026.0349.5