Selective Test-Time Reasoning with Relation-Aware Video Evidence for VRR-QA

VRR-QA asks models to recover spatial and temporal relations from video, including cases where viewpoint, motion, or occlusion makes a single frame ambiguous. We address these questions with Adaptive Dense Evidence Refinement, an inference pipeline built around a direct video-language prediction. Alternative views produce candidate answers and help select questions for further inspection. For a selected question, the system gathers timestamped observations, probes the relevant relation, and compares the initial answer with a candidate. It changes the answer when the observations support the candidate, contradict the initial prediction, and pass the ambiguity checks. The complete system achieves 90.07 average accuracy and 87.81 macro average accuracy on the test set; direct Gemini prompting achieves 70.47 average accuracy. The procedure uses pretrained models without task-specific weight updates.