MetaphorVU: Towards Metaphorical Video Understanding

Researchers introduced MetaphorVU-Bench to evaluate how well multimodal large language models understand metaphorical videos.

They also proposed MetaphorBoost, a knowledge-graph-based method that improves inference-time performance on this task.

The work highlights a major weakness in current multimodal AI and offers a more human-like way to reason about figurative visual content.