Long-horizon cognition
We study long-horizon understanding through full-length films: connecting scattered images and dialogue, understanding characters and events, and revising earlier interpretations as later scenes unfold. Current work includes film indexing, cross-scene retrieval, and a review workspace.
Across thousands of moments, how should an agent decide where to look, what to remember, and when to reinterpret?
A later scene can change our understanding of a character’s motives or turn an ordinary earlier image into a crucial clue. We study how systems maintain a current interpretation, acquire information in response to a question, and revise that interpretation when new information arrives.
From local moments to complete works.
Connect images, dialogue, and information across scenes. Organize relationships between characters and events to provide context for question-driven understanding.
Let the task guide observation.
Explore which information to acquire under a limited observation budget, when to look closer or revisit, and how later events should prompt revisions to earlier interpretations.
Discuss and revise in a shared workspace.
Connect source-media navigation, candidate interpretations, and human review so that understanding can be examined, discussed, and refined.