Annotations as Rollouts: Efficient and Scalable Reinforcement Learning for Video MLLMs

Benchmark Model Rank Results
video-question-answering-on-videoholmesVideo-ORA-9B#1Accuracy: 65.5
temporal-sentence-grounding-on-activitynet-timelensVideo-ORA-9B#2mIoU: 63.6R@1 IoU=0.3: 77.4R@1 IoU=0.5: 69.5R@1 IoU=0.7: 53.3
temporal-sentence-grounding-on-charades-timelensVideo-ORA-9B#2mIoU: 61.8R1@0.5: 71.6R1@0.7: 42.4R1@0.3: 84.5
referring-video-object-segmentation-on-mevisVideo-ORA-9B#3J&F: 61.3
temporal-sentence-grounding-on-qvhighlights-timelens-validationVideo-ORA-9B#3mIoU: 72.5R@1 IoU=0.3: 84.9R@1 IoU=0.5: 77.6R@1 IoU=0.7: 64.0
referring-video-object-segmentation-on-reasonvosVideo-ORA-4B#4J&F: 63.8
video-question-answering-on-mmvuVideo-ORA-9B#6Accuracy: 77.0
zero-shot-video-question-answer-on-video-mme-w-o-subsVideo-ORA-9B#10Accuracy (%): 76.7
video-question-answering-on-mvbenchVideo-ORA-9B#20Avg.: 72.5