Long-Video Understanding

Long-video understanding studies how models can answer questions, track entities, and reason about events spread across minutes or hours of video. Common approaches select or compress frames, build hierarchical visual or event memories, and use retrieval tools or agents to find and verify relevant evidence; some also combine audio, speech, and video. Benchmarks such as Video-MME, LongVideoBench, LVBench, and MLVU test temporal grounding, long-range recall, and multi-event reasoning, and results show that strong short-clip performance does not guarantee reliable long-video understanding. A central tension is how to reduce computation without losing sparse or distant evidence, while improving grounding, temporal consistency, and the ability to recognize when evidence is insufficient.