Agentic video reasoning systems gather and assess video evidence across multiple stages, selecting relevant spans as visual evidence for answering questions. According to a new arXiv paper, these systems face a challenge: when new candidates or observations appear, they prompt a reassessment of the current span. The paper, titled VideoSlate, introduces a stateful commit control mechanism to address this issue.
The abstract is truncated, but the title and opening indicate that VideoSlate aims to provide a more disciplined way to decide when to commit to a selected span, rather than constantly revisiting it. The approach is described as "stateful," implying that the system maintains a record of what has been committed and uses that state to control future reassessments. This could reduce unnecessary re-evaluation and improve temporal grounding in agentic video reasoning tasks.