MVBench: Results as Published

MVBench tests temporal understanding of video. It has 20 tasks that cannot be solved from a single frame, made by turning static image tasks into dynamic ones, such as action sequence, moving direction, and counterfactual inference. Public video annotations are converted into multiple-choice questions, 200 per task, so the correct option is the annotated ground truth. The score is accuracy, and the paper also reports the average accuracy across tasks.

Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.