Visual prompt engineering for video models

A new arXiv paper introduces VIPE, or visual prompt engineering, for improving video model reasoning.

The method edits input images to make tasks easier for models, rather than changing the model itself.

Researchers report gains across multiple visual reasoning tasks, sometimes outperforming text prompting and extra test-time compute.

The approach could offer a low-cost way to boost foundation video models without larger models or heavier inference.