Discriminative methods that have Vid-LLMs directly judge frame relevance suffer from an Interpretation–Selection Gap, where the model understands the query but may not accurately execute the selection action.