Paper Overview
- Field: Computer Vision (CV)
- Authors: Shijie Wang, Honglu Zhou, Ziyang Wang, Ran Xu, Caiming Xiong
- Published: 2026-07-13
- arXiv: 2607.11862
- ST-Evidence benchmark: the first human-verified benchmark for both discriminative and generative pixel-level grounding.
- Key finding: evaluations of state-of-the-art models reveal a critical decoupling between QA accuracy and genuine visual perception—one that cannot be closed by scaling alone.
- ST-Evidence-Instruct dataset: a scalable automated generation pipeline producing 160k training samples.
- Strong results: fine-tuning grounding Video LLMs on this dataset yields large improvements over the same-sized UniPixel baseline (t-mean +27.2, J&F +13.8 on 7B models), establishing a robust baseline for explainable, evidence-based video understanding.
Summary
Current Video Large Language Models (Video LLMs) excel in question answering (QA) but largely operate as black boxes, providing textual answers without verifiable visual grounding. Existing explainability efforts rely on textual rationales or sparse bounding boxes, which struggle to capture complex video dynamics such as occlusions and non-rigid deformations.
The paper proposes Evidence-Backed Video Question Answering (E-VQA), a novel task requiring models to jointly output a semantic answer and precise spatio-temporal evidence: temporal segments and dense, tracked object segmentation masklets.
Key Contributions
*Auto-collected on 2026-07-15.*