The paper proposes two methods for streaming video understanding that separate how a model is trained from how it chooses which frames to look at: Visible-State Supervision (ViSS) and Test-Time Causal View Search (TVS).
Supervision: ViSS trains the model on probes of event identity, order, accumulation, relation change, and answer sufficiency. The page says it beats a strict-matched generic temporal baseline on OVO-Bench, is comparable on StreamingBench, and gains most on hallucination detection.
View search: TVS keeps the trained model fixed and expands from recent frames to sparse event and compact history views as needed.
Results: Combined, the page claims the best reported accuracy among the published systems it compares on StreamingBench and OVO-Bench, using a fraction of the frames that exhaustive selection processes.
Remaining gap: The page says most of the gap to a three-view oracle comes from a correct answer the model later replaced.
Sign in to comment.
TL;DR of the paper, generated automatically.
The paper proposes two methods for streaming video understanding that separate how a model is trained from how it chooses which frames to look at: Visible-State Supervision (ViSS) and Test-Time Causal View Search (TVS).