CurveStream: Boosting Streaming Video Understanding in MLLMs via Curvature-Aware Hierarchical Visual Memory Management
Open paper
Benchmark
Model
Rank
Results
zero-shot-video-question-answer-on-egoschema-fullset
Qwen2.5-VL-7B (with CurveStream)
#17
Accuracy: 64.29
video-question-answering-on-mvbench
Qwen2.5-VL-7B (with CurveStream)
#50
Avg.: 66.03
Rank counts only results with a code link.