Supervising the Visible State: Separating Representation from View Selection in Streaming Video
The combined method reports the highest accuracy among the published systems compared on two streaming-video benchmarks, while processing far fewer frames than exhaustive selection.
Visible-State Supervision trains models to distinguish event identity, order, accumulation and changing relations—details ordinary action labels can leave unconstrained. Test-Time Causal View Search keeps the trained model fixed and expands access from recent frames to event and history views as needed. Supervision particularly improves hallucination detection and causal reasoning; evidence access particularly improves event recall and forward reasoning. Replay analysis finds that most of the remaining gap to a three-view oracle comes from correct answers being replaced later.
HF Daily Papers' note
Visible-State Supervision trains models to distinguish event identity, order, accumulation and changing relations—details ordinary action labels can leave unconstrained. Test-Time Causal View Search keeps the trained model fixed and expands access from recent frames to event and history views as needed. Supervision particularly improves hallucination detection and causal reasoning; evidence access particularly improves event recall and forward reasoning. Replay analysis finds that most of the remaining gap to a three-view oracle comes from correct answers being replaced later.
HF Daily Papers' note
score 4