COVER Gives Video Grounders Statistical Coverage Guarantees

COVER wraps any video-grounding model with statistical coverage guarantees and skips retraining the model underneath it, per the paper posted at arXiv 2608.07434.

That matters because the benchmarks these models get scored on may not be measuring what they claim. When the paper's authors had independent annotators re-label the same query-video pairs, the new labels overlapped with the originals by less than half. That gap showed up on a large fraction of samples. Ground truth for a video moment, in other words, is closer to a range of plausible answers than a single fixed interval.

COVER doesn't try to fix the labels themselves. It calibrates a temporal nonconformity quantile on a held-out set, then widens each model's predicted window until it hits a target coverage rate. The guarantee holds for any finite sample, as long as the data is exchangeable. Across three benchmarks and five different grounding models, it held that target coverage rate every time, without touching the underlying model's weights.

The paper doesn't say what happens once that exchangeability assumption breaks. That's the likeliest failure mode for a model deployed on video it wasn't calibrated against.