I Seek You in Videos: Identity-Conditioned Queries for Person-Centric Video Reasoning
The paper introduces a benchmark and training setup for finding and reasoning about a specific person across video using a reference image.
Its ICQ task asks models to connect a video with an image of the target person, then answer questions about identity, behavior, and time. The authors release ISYV-Bench with 1,377 real-world videos and QA pairs across six difficulty levels, plus a 75K-sample training set. They report that mainstream closed and open multimodal LLMs struggle most with cross-domain identity matching and long-horizon tracking. Their ISYV-Model beats strong baselines and partly approaches closed-source performance. ArXiv · AI/CL/LG's note
Its ICQ task asks models to connect a video with an image of the target person, then answer questions about identity, behavior, and time. The authors release ISYV-Bench with 1,377 real-world videos and QA pairs across six difficulty levels, plus a 75K-sample training set. They report that mainstream closed and open multimodal LLMs struggle most with cross-domain identity matching and long-horizon tracking. Their ISYV-Model beats strong baselines and partly approaches closed-source performance. ArXiv · AI/CL/LG's note
score 4