Sparse-frame video-language models, which answer questions about videos using only a few sampled frames, can change their answer when the sampling timestamps are shifted, according to a new arXiv preprint. The authors report that this instability is exposed by raw agreement between runs, but that simple agreement is not a better error-ranking signal than token confidence.
The study also finds that raw agreement does not indicate which additional frames a model should reobserve to improve its answer. The abstract does not describe the proposed method in detail, but it frames the problem as one of learning when to trust a model's output and when to request more information.
Because only the abstract is available, the specific datasets, model architectures, and evaluation results are not covered here. The authors position the work as a step toward making sparse-frame VideoQA more reliable under temporal sampling perturbations.