The abstract opens by noting that while collecting data for model training is relatively easy, curating high-quality datasets from large collections is inherently challenging. Coreset estimation methods address this problem by identifying the best data subsets to use for training, but many existing approaches rely on task-specific supervision or labels.
The new paper, titled "Diversity Without Labels: Task-Agnostic Coreset Selection Across Perception and Action," proposes a coreset selection method that is explicitly task-agnostic and label-free. The emphasis on diversity suggests the method selects subsets that represent the full range of data variation, without needing annotated examples to guide the selection.
The abstract indicates the approach is designed to apply across both perception and action tasks, though the released text is truncated before giving details on the method or experimental results. As such, the core claim is that meaningful diversity can be captured without labels, potentially reducing the annotation burden in dataset curation.