ViSS: Improving Video State Supervision and Hallucination Detection
October 25, 2026
Visible-State Supervision (ViSS) trains video models on source-grounded probes of event identity, order, and relation changes to prevent state under-specification. This approach improves performance on OVO-Bench and enhances a model's ability to detect hallucinations in streaming video.
HOW THIS AFFECTS YOU
●
researcherThis addresses the issue of models ignoring temporal order and accumulation in video reasoning.