Puzzle Curriculum GRPO for Vision-Centric Reasoning

Benchmark Model Rank Results
video-question-answering-on-tempcompass-overallQwen2.5-VL-7B (Jigsaw)#3Accuracy: 75.4
mathematical-reasoning-on-mathverseQwen2.5-VL-7B (Mix)#12Accuracy: 58.1
video-question-answering-on-videommmuQwen2.5-VL-7B (Mix)#13Accuracy: 42.66
multimodal-reasoning-on-mathvisionQwen2.5-VL-7B (Jigsaw)#47Accuracy: 30.92
vision-language-understanding-on-mmstarQwen2.5-VL-7B (Mix)#58Accuracy: 65.2
multimodal-reasoning-on-mathvistaQwen2.5-VL-7B (Rotation)#60Accuracy: 71.7
video-question-answering-on-mvbenchQwen2.5-VL-7B (Mix)#64Avg.: 61.0