CutVerse: A Compositional GUI Agents Benchmark for Media Post-Production Editing

Researchers introduced CutVerse, a benchmark for evaluating autonomous GUI agents on realistic media post-production editing tasks.

The benchmark includes expert demonstrations from seven professional applications and 186 editing tasks, with a parser that turns screen recordings and interaction logs into structured trajectories.

In evaluations, existing agents completed only 36.0% of tasks, showing they still struggle with complex, long-horizon creative workflows.

The results highlight a major gap between current GUI agents and the demands of real-world media editing in tools like Premiere Pro and Photoshop.