• 강사 : 김현우 교수 (KAIST 전산학부)
• 일시 : 2025. 10. 20. (월) 16:00~17:30

This talk explores the frontier of deep video understanding through the lens of foundation models, aiming to unify perception, reasoning, and action in video agents. It begins by revisiting Video Question Answering, highlighting how to overcome modality bias and improve cross-modal alignment between vision and language. Next, this talk introduces a light-weight token reduction method and a new video foundation model of efficient long video understanding. Moving beyond passive understanding, this talk will briefly discuss Vision-Language-Action (VLA) models that bridge perception with decision-making, enabling foundation models to act upon their visual understanding.