ResearchTasksHeld-out GeneralizationFollowBenchmarksDataset NameSOTA MethodSortMost resultsRecently updatedMost papersApplyDataset NameSOTA methodMetricTrendResultsLast UpdatedBASALT Top 10% hardest (n=30 clips)PROWL0.903Latency (s)3May 20, 2026BASALT Top 25% hardest (n=75 clips)PROWL0.81Latency3May 20, 2026BASALT Mean (n=300 human clips (100/task))PROWL0.629Latency3May 20, 2026