arXivPaperNeeds Review
Evidence-Backed Video Question Answering
E-VQA requires video answers to include tracked pixel-level evidence, revealing when good QA scores hide weak perception and supplying grounded training data.
arXiv
Source Summary
**E-VQA** requires a semantic answer plus temporal segments and dense tracked segmentation masks. Its human-verified ST-Evidence benchmark finds that QA accuracy can diverge from visual grounding, a gap the authors say scaling alone does not close.
Practical Implication
For video agents, evaluate whether the cited object persists through motion, occlusion, and deformation rather than scoring text alone. The accompanying **160k-scale ST-Evidence-Instruct** dataset provides training examples that tie reasoning to visible evidence.
Agent-Ready Context
**E-VQA** requires a semantic answer plus temporal segments and dense tracked segmentation masks. Its human-verified ST-Evidence benchmark finds that QA accuracy can diverge from visual grounding, a gap the authors say scaling alone does not close. For video agents, evaluate whether the cited object persists through motion, occlusion, and deformation rather than scoring text alone. The accompanying **160k-scale ST-Evidence-Instruct** dataset provides training examples that tie reasoning to visible evidence. On a **7B model**, fine-tuning beats a size-matched UniPixel baseline by **+27.2 t-mean and +13.8 J&F**. Those gains are specific to the reported setup and grounding metrics; the material does not establish equivalent gains for downstream video-agent tasks.
Context Map
benchmarkvideo#generative-media#agent-evals#benchmark-integrityUncertainty
On a **7B model**, fine-tuning beats a size-matched UniPixel baseline by **+27.2 t-mean and +13.8 J&F**. Those gains are specific to the reported setup and grounding metrics; the material does not establish equivalent gains for downstream video-agent tasks.