StreamPro-Bench: Benchmarking Proactive Video Understanding
StreamPro-Bench, a novel benchmark, assesses the capability of AI models to actively comprehend streaming video by requiring them to determine when to react based on incomplete observations. Traditional benchmarks adhere to a 'see-then-answer' approach, which limits proactive reasoning to a lag in perception. StreamPro-Bench tackles the significant disparity between silence and response signals in streaming data, emphasizing the importance of optimizing both the accuracy and timing of responses. This benchmark is designed to enhance proactive decision-making in the realm of streaming video understanding.
Key facts
- StreamPro-Bench is a new benchmark for proactive streaming video understanding.
- Existing benchmarks follow a 'see-then-answer' paradigm.
- Proactive models must balance early prediction against sufficient evidence.
- Training proactive models is challenging due to imbalance between silence and response signals.
- The benchmark requires models to decide when to respond, not just what to respond.
- StreamPro-Bench addresses joint optimization of response correctness and timing.
- The benchmark is introduced in arXiv:2605.16381.
- Proactive reasoning is distinguished from delayed perception.
Entities
Institutions
- arXiv