Visual Prompt Engineering for Video Models

Researchers introduced VIPE, or visual prompt engineering, which edits input images to make video reasoning tasks easier for models.

The method improves performance across multiple tasks and can outperform classic text-based prompt engineering.

The authors also say VIPE can beat test-time scaling in some cases, offering a low-compute way to boost reasoning.

The approach works by using an image editing model to alter the visual prompt rather than changing the text instruction.