VESTA is a training-free agent for long-video understanding that replaces the usual single retrieval behavior with a policy-steered approach. Before searching, an intent router decides whether the question needs focused, broad-coverage, or contrastive evidence, and also configures how evidence will be accounted for during exploration. This lets the agent acquire the right kind of evidence from the start, while still allowing the reasoner to freely verify, re-query, or inspect outside the retrieved set.
On Video-MME-v2, VESTA improves average accuracy by 2.7 points over VideoARM and gains on all six reported metrics. It also improves by 6.9 points on the LongVideoBench long subset and 1.5 on LVBench, and matches VideoARM on EgoSchema. The gains suggest that matching retrieval strategy to question type is a simple but effective way to avoid early failures in long-video reasoning.