Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning

Benchmark Model Rank Results
moment-retrieval-on-charades-staVideoReasoner#8R@1 IoU=0.5: 70.4R@1 IoU=0.7: 45.7R@1 IoU=0.3: 81.9mIoU: 59.1
video-question-answering-on-tomatoVideoReasoner (Qwen2-VL-7B-Base + RL)#12Accuracy: 27.8
video-question-answering-on-mmvuVideoReasoner (Qwen2.5-VL-7B-Instruct + RL)#18Accuracy: 68.0
zero-shot-video-question-answer-on-video-mme-w-o-subsVideoReasoner#42Accuracy (%): 62.0
video-question-answering-on-lvbenchVideoReasoner#45Accuracy: 44.6
video-question-answering-on-mlvuVideoReasoner#50Accuracy: 64.6