Source-Free Domain Adaptation (SFDA) aims to transfer a model pretrained on a source domain to an unlabeled target domain without accessing the original source data.
Early single‑model approaches rely on self‑refinement, which easily suffers from confirmation bias and struggles to correct systematic errors. Recent works introduce Vision‑Language (ViL) models as external knowledge sources, but most adopt a unidirectional paradigm where the ViL model only supervises the source‑pretrained model, overlooking a structural property: the two models often have complementary failure modes—when one predicts incorrectly, the other may be correct, offering a natural chance for mutual correction. Without ground‑truth labels, identifying the reliable model for each sample is non‑trivial, and naïvely swapping predictions can propagate errors.
To address this, we propose SafeCut. The key idea is to use the cut statistic as a label‑free reliability measure to gate cross‑model supervision. For each sample we compute the cut values of both models; a smaller cut indicates higher confidence. Based on the relative reliability we dynamically decide the supervision direction and adjust its strength, allowing only the more reliable model to provide soft labels to the less reliable one. This per‑sample gating amplifies true corrections while suppressing miscorrections.
We also provide theoretical analysis showing that the reliability‑gated mechanism yields a net‑positive expected correction signal, guaranteeing overall performance improvement. Extensive experiments on diverse SFDA benchmarks demonstrate that SafeCut consistently outperforms existing methods, confirming the effectiveness of safeguarding mutual correction via cut statistics.
Review