StreamPro: From Reactive Perception to Proactive Decision-Making in Streaming Video

Benchmark Model Rank Results
video-question-answering-on-ovo-benchStreamPro-GRPO (3B)–Real-Time Visual Perception: 62.5Backward Tracing: 34.9
video-question-answering-on-ovo-benchStreamPro-GRPO (4B)–Real-Time Visual Perception: 66.9Backward Tracing: 45.2
video-question-answering-on-ovo-benchStreamPro-SFT (3B)–Real-Time Visual Perception: 62.0Backward Tracing: 40.0
video-question-answering-on-ovo-benchStreamPro-SFT (4B)–Real-Time Visual Perception: 67.9Backward Tracing: 46.0
video-question-answering-on-streamingbenchStreamPro-GRPO (3B)–Real-Time Visual Understanding: 76.0
video-question-answering-on-streamingbenchStreamPro-GRPO (4B)–Real-Time Visual Understanding: 78.9
video-question-answering-on-streamingbenchStreamPro-SFT (3B)–Real-Time Visual Understanding: 77.1
video-question-answering-on-streamingbenchStreamPro-SFT (4B)–Real-Time Visual Understanding: 79.3