MLVU

MLVU (Multi-task Long Video Understanding) tests long video comprehension: videos of up to two hours, nine different task types (plot questions, needle-in-video search, anomaly detection and more). The benchmark shows whether a model truly follows a long video or only spots single frames.

MLVU measures understanding videos instead of static images. In detail, it checks:

• long videos instead of short clips
• nine different comprehension task types
• temporal reasoning: what happened and when
• search of specific visual details within long videos

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better.

What is the MLVU benchmark?

MLVU (Multi-task Long Video Understanding) tests long video comprehension: videos of up to two hours, nine different task types (plot questions, needle-in-video search, anomaly detection and more). The benchmark shows whether a model truly follows a long video or only spots single frames.

What does MLVU test?

MLVU mainly measures understanding videos instead of static images. Concretely, it checks: long videos instead of short clips; nine different comprehension task types; temporal reasoning: what happened and when.

How is the MLVU score calculated?

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better. The declared maximum score of this benchmark is 100.0.

Where do I get MLVU data and scores?

The official benchmark page/reference is linked on this page. Every score on Open LLM Radar is also available as open data: download the scores JSON (http://open-llm-radar-api:8000/api/v1/benchmarks/mlvu/scores) or pull it directly from the Open LLM Radar API with any HTTP client.