The authors created EvaOmni, a benchmark of 392 videos and 1,027 question-answer pairs, each annotated with audio and visual evidence intervals by experts.
Evidence-grounded benchmark forces models to show their reasoning for audio-visual tasks
EvaOmni requires models to localize audio and visual evidence, uncovering that even top performers like Gemini3.7-Flash have weak temporal grounding.
Chinese Tech
Xiaoxing Hu · Wenxi Yao · Chaoyuan Shi · Ziqi Ye · Xiaorong Guo · Peixuan Li · +14 more
Shanghai Jiao Tong University · Fudan University · Beijing Institute of Technology · Tongji University · The Hong Kong University of Science and Technology (Guangzhou)
Research Digest··2 min read
The authors introduce EvaOmni, a benchmark of 1,027 audio-visual question-answer pairs requiring models to provide supporting evidence intervals.
§