TimeScope: Video Large Multimodal Models Struggle with Temporal Understanding

AI Impact Summary

The TimeScope benchmark reveals a significant gap between the claims of large multimodal models regarding their ability to process long videos and their actual performance. Specifically, models struggle with tasks requiring true temporal comprehension, often relying on surface-level retrieval rather than synthesizing information or perceiving fine-grained motion across extended sequences. This highlights a critical area for improvement in vision-language models, particularly as they are increasingly deployed in applications requiring understanding of dynamic, long-form content.