Hugging Face Daily PapersYibo Ma, Qianqian Zhang, Peng Liu1 min readpaperadvanced
TRACE: Temporal Audit and Condition-aware Evaluation of Streaming Video Understanding
Summary
TRACE is a condition-aware benchmark for streaming video understanding that makes evidence timing and trigger conditions explicit, and measures answer quality, timeliness, workload, and reliability. Experiments on 1,240 records show that identical QA accuracy can hide large differences in completion, false alarms, and processing cost.
- TRACE adds evidence‑timing and instruction‑dependent trigger annotations to streaming video tasks.
- The Core‑Adapter protocol controls information flow while logging history processing and response events.
- Metrics go beyond accuracy: they include timeliness, workload, false‑alarm rate, missed windows, and reliability.
- Eight public models achieve similar QA scores but differ widely in completion rates and generation cost.
Anyone building or researching streaming video QA systems should care, because single‑score evaluations miss critical operational trade‑offs.
7/10/)