Less from More: Reinforcing Sparse Video Reasoning from Dense References

Benchmark Model Rank Results
video-question-answering-on-videommmuSAVER-4B (0.1 fps) (extra data)#2Accuracy: 72.7
video-question-answering-on-mmvuSAVER-4B (extra data)#13Accuracy: 70.9
zero-shot-video-question-answer-on-video-mme-w-o-subsSAVER-4B (extra data)#20Accuracy (%): 71.8
video-question-answering-on-mvbenchSAVER-4B (extra data)#52Avg.: 65.3
moment-retrieval-on-charades-staSAVER-4B (extra data)#79mIoU: 54.8