DataVista: Diagnosing Multimodal LLMs on Data Video Understanding
DataVista finds leading multimodal models still well short of expert-level understanding on animated data videos.
The benchmark covers 961 real-world data videos and 6,775 questions across data perception, temporal reasoning, and narrative understanding. In tests of 19 MLLMs, Gemini-3.1-Pro led with 70.0% accuracy. The hardest areas were causal reasoning and narrative structure. More frames and subtitles helped with reading and timing tasks, but did little for narrative understanding. HF Daily Papers' note
The benchmark covers 961 real-world data videos and 6,775 questions across data perception, temporal reasoning, and narrative understanding. In tests of 19 MLLMs, Gemini-3.1-Pro led with 70.0% accuracy. The hardest areas were causal reasoning and narrative structure. More frames and subtitles helped with reading and timing tasks, but did little for narrative understanding. HF Daily Papers' note
score 5