MVU-Eval: A Multi-Video Understanding Benchmark for MLLMs

MVU-Eval is a benchmark dataset for evaluating Multimodal Large Language Models on multi-video understanding tasks. It was created by the MVU-Eval-Team and last updated on Hugging Face in November 2025. The dataset addresses a gap in existing single-video benchmarks by focusing on scenarios requiring analysis across multiple videos, such as sports analytics and autonomous driving.

Use Cases

• Benchmarking MLLM performance on multi-video reasoning tasks based on the described focus.
• Training models for sports analytics based on the multi-video scenario mentioned in the description.
• Developing autonomous driving perception systems based on the multi-video understanding need cited.
• Researching cross-video temporal and causal inference based on the dataset's stated purpose.

Strengths

• Designed as the first comprehensive benchmark for multi-video understanding, addressing a stated gap.
• Focuses on real-world application scenarios explicitly mentioned, such as sports analytics and autonomous driving.
• Last updated on 2025-11-12, indicating recent maintenance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and dataset size are unknown, which may limit suitability assessment.
• The full description is hosted externally, requiring a click-through for complete details.

Provenance

Source

MVU-Eval-Team on Hugging Face

Freshness

Last updated 2025-11-12 02:29:28