What it is
Introduces StreamArena, an hour-scale open-ended benchmark for streaming video agents, and StreamMind, a two-tier agent that separates fast interaction from persistent multimodal memory. Instead of short MCQ clips or single-path streaming memory, it benchmarks hour-scale open-ended causal QA and uses a two-tier agent: frontend workers handle low-latency interaction/monitoring, backend workers build persistent multimodal memory and do recall/search.
Key results
- 243 videos, 3,646 open-ended QA pairs; avg video length 88.8 min
- StreamMind beats strongest streaming baselines by 58.4% RTP and 53.7% HR
- Cuts pooled query-to-answer latency by 66.2% with the same Qwen3.5-397B backbone
Shows streaming agents need decoupled real-time interaction and persistent multimodal memory to handle long-horizon recall, proactive alerts, and tool use together.