Program-Guided Temporal Reasoning over Routed Video Evidence for TimeLogicQA

A temporal video question may turn on a brief transition, a repeated action, or the boundary between two events. Temporal Evidence Routing approaches these questions by first finding the observations needed to check the requested relation. It parses the events, options, and temporal condition, then supplies ordered frames for short clips or event-focused windows for longer videos. A multimodal model records event occurrences and their intervals. Temporal checks use these records to produce a candidate answer, which is combined with a base prediction through evidence and confidence checks. The system operates with fixed pretrained weights. On TimeLogicQA, it achieves 81.8 AvgAcc against 62.50 for direct Gemini prompting, a gain of 19.30 percentage points.