VAST: Video Ability-Stratified Taxonomy for Data-Efficient Video Reasoning

Benchmark Model Rank Results
spatial-reasoning-on-vsi-benchVideoVAST-7B (extra data)–Accuracy: 33.5
video-question-answering-on-mvbenchVideoVAST-7B (extra data)–Avg.: 66.3
video-question-answering-on-tempcompass-overallVideoVAST-7B (extra data)–Accuracy: 74.2
video-question-answering-on-videoholmesVideoVAST-7B (extra data)–Accuracy: 42.5
video-question-answering-on-videommmuVideoVAST-7B (extra data)–Accuracy: 50.6