StreamPro AI shifts video models from reactive to proactive reasoning, scoring 4x better

Current video understanding models follow a passive “see-then-answer” paradigm: they wait for clear evidence before responding. This reduces proactive reasoning to delayed perception. The new paper “StreamPro: From Reactive Perception to Proactive Decision-Making in Streaming Video” by Ao Li and 9 co-authors tackles the harder problem of deciding when to speak, not just what to say. The authors introduce StreamPro-Bench, a benchmark that scores models on three axes: Perception Understanding, Temporal Reasoning, and Proactive Agency. The last measures a model’s ability to make early, yet reliable, decisions from incomplete streams—a critical skill for real-time applications like live surveillance, autonomous driving, or interactive assistants.

To train such models, the team proposes a two-stage framework also called StreamPro. First, they use CB-Stream Loss during supervised fine-tuning to mitigate the extreme imbalance between long periods of irrelevant silence and short, critical response windows. Second, they apply Group Relative Policy Optimization (GRPO) with a multi-grained reward design that penalizes both wrong answers and poor timing—optimizing correctness and decision delays jointly. Results show dramatic gains: StreamPro achieves 41.5 on its proactive benchmark, far exceeding the previous best of 10.4, while maintaining strong real-time performance (78.9 on StreamingBench-RTVU). The work signals a shift from passive video understanding to agents that can proactively engage with streaming video.