Reinforcement Learning with Verifiable Rewards (RLVR) offers a natural framework for adapting pretrained models to video temporal grounding, where generated intervals can be scored directly against ground‑truth intervals.
Existing overlap verifiers typically score each rollout independently, leaving the joint structure of the rollout group untapped.
We introduce SUTURE, which conditions verification on the entire rollout group and exploits its structure at two complementary scales: disagreement across rollouts controls how strongly the target is reweighted, while coverage at each position determines where reward mass is redistributed.
We prove that the resulting verifier admits an exact decomposition into the standard IoU term plus a covariance correction determined by the rollout group.
A local gradient diagnostic reveals a preference for responses covering relatively less supported target regions within the analyzed groups.
Across five temporal grounding benchmarks, SUTURE improves grounding performance at every reported IoU threshold. Its trained policy also exhibits less video‑start anchoring in reasoning traces: for later events, the first temporal mention more often overlaps the annotated target.
Together, these findings demonstrate that the joint structure of a rollout group can provide a more informative temporal verifier. Review