E-VQA introduces spatio-temporal evidence for video QA
July 13, 2026
The E-VQA task and ST-Evidence benchmark require Video LLMs to provide semantic answers alongside precise temporal segments and dense object segmentation masklets. This addresses the lack of verifiable visual grounding in existing black-box models.
HOW THIS AFFECTS YOU
●
builderUtilize spatio-temporal masks to build more explainable video-based AI features.
●
researcherMove toward pixel-level grounding evaluation for video models.