VideoKR: Towards Knowledge- and Reasoning-Intensive Video Understanding

Benchmark Model Rank Results
video-question-answering-on-videommmuVideoKR (SFT + RL) on Qwen3-VL-8B-Instruct (Frames=128)#6Accuracy: 63.0
video-question-answering-on-mmvuVideoKR (SFT + RL) on Qwen3-VL-8B-Instruct (Frames=128)#22Accuracy: 64.8
video-question-answering-on-mvbenchVideoKR (SFT + RL) on Qwen2.5-VL-7B-Instruct (Frames=128)#38Avg.: 68.9