Fine-tuned vision models often lose accuracy when faced with distribution shifts. A new arXiv preprint proposes Barycentric Anchored Search, a post-hoc procedure that operates directly in weight space to improve such models.

The method requires no additional gradient training and does not increase inference cost, making it a lightweight fix. According to the abstract, it improves fine-tuned vision models under distribution shift by anchoring them to a pre-trained base.

The paper focuses on vision models and reports improvements under distribution shift. As a single source, we can only report the authors' claims; no independent verification is provided in the abstract.