MVEB Benchmarks 33 Video Embedding Models Across 23 Tasks, No Single Winner

MVEB evaluates 33 models across classification, retrieval, clustering, and QA tasks, finding MLLM-based embeddings lead on classification and QA while multimodal binding leads on retrieval. Audio improves performance by up to 6 points when labels were derived from both modalities but hurts when labels are vision-only.