{"ID":6023320,"CreatedAt":"2026-07-08T01:00:23.257252134Z","UpdatedAt":"2026-07-10T01:11:38.759438437Z","DeletedAt":null,"paper_url":"https://arxiv.org/abs/2607.05726","arxiv_id":"2607.05726","title":"Association Restoration Test: Revealing Restorable Shortcuts after Unlearning","abstract":"Association unlearning aims to disable learned label-attribute shortcuts while preserving task performance. Existing evaluations mainly measure output-level robustness or probe whether shortcut attributes remain readable in frozen features, but neither test determines whether a retained association remains functionally usable by the original classifier. We propose the Association Restoration Test (ART), a post-hoc diagnostic for functional shortcut restorability. ART estimates class-conditional association directions, amplifies residual components, and evaluates the modified features with the original classifier head. Across Waterbirds, CelebA, SpuCoDogs, and an ISIC timestamp-artifact extension, we show that output metrics, representation probes, and ART characterize distinct aspects of shortcut mitigation. These findings motivate restoration-aware evaluation for unlearning and shortcut-mitigation methods that target learned associations rather than individual classes or concepts.","short_abstract":"Association unlearning aims to disable learned label-attribute shortcuts while preserving task performance. Existing evaluations mainly measure output-level robustness or probe whether shortcut attributes remain readable in frozen features, but neither test determines whether a retained association remains functionally...","url_abs":"https://arxiv.org/abs/2607.05726","url_pdf":"https://arxiv.org/pdf/2607.05726v1","authors":"[\"Amy Lu\",\"Changxiu Ji\"]","published":"2026-07-07T01:20:52Z","proceeding":"cs.CV","tasks":"[\"cs.CV\",\"cs.LG\"]","methods":"[]","has_code":false}