VideoSeeker: Incentivizing Instance-level Video Understanding via Native Agentic Tool Invocation

Researchers introduced VideoSeeker, a framework for instance-level video understanding that combines visual prompts with agentic tool use.

It uses a fully automated data synthesis pipeline and trains with both supervision and reinforcement learning.

The system improves fine-grained spatiotemporal localization and video retrieval, especially for precise object and event grounding.

Reported results show gains over strong baselines and even closed-source models such as GPT-4o and Gemini-2.5-Pro.