VideoVista: Benchmarking Diverse and Complex Video-Language Interaction for MLLMs

Multimodal AI's video comprehension reflects real-world dynamic understanding, but mainstream Video-Language Interaction (Video QA) benchmarks are primarily English-centric, narrow-domain, and only test basic perceptual or reasoning skills, lacking holistic assessment of model performance across diverse languages, domains, and complex spatiotemporal scenarios. To address these gaps, we present VideoVista, a versatile multilingual and multidomain video comprehension benchmark, built via scalable, quality-controlled automatic QA generation and manual refinement. Evaluations on 41 leading MLLMs reveal three key findings: 1) Small 7/8B open-source models rival large commercial ones with Chain-of-Thought (CoT) reasoning; 2) Models perform poorly (accuracy $\lt $63%) on tasks requiring complex temporal reasoning (e.g., Streaming QA), with most open-source models degrading further when key information appears late in videos; 3) Fine-grained tasks (e.g., anomaly detection) see sharp performance drops and models exhibit cultural bias. We release the benchmark, automatic QA generation framework, and associated pretraining, instruction-tuning, and reinforcement learning datasets at https://github.com/HITsz-TMG/VideoVista.