A research paper proposes a new method for making AI classifiers hold up when the data they're tested on looks nothing like the data they trained on.
The framework, called SFT+RL, tackles a known weak spot in unsupervised domain adaptation: models that perform well on clean source data tend to fall apart when the target data has been adversarially perturbed. The approach runs in two stages. First, it adversarially fine-tunes a linear classifier on top of CLIP's visual encoder using a technique called PGD-based perturbation, partially unfreezing CLIP's projection layer to let the model adapt without losing its general visual understanding. Second, a reinforcement learning stage progressively assigns labels to unlabeled target samples, filtering them through a decaying confidence threshold to keep only the high-quality guesses before training on a mix of clean and adversarial examples.
The results are hard to dismiss: across three standard benchmark datasets — OfficeHome, PACS, and VisDA — SFT+RL averaged a 10.2% gain in clean accuracy and a 15.8% gain in adversarial robustness over existing methods. That kind of gap matters because real-world deployments rarely get clean, in-distribution data, and current approaches force an uncomfortable trade-off between the two.
Building on CLIP rather than training from scratch is a sensible move, but it also means the method inherits whatever biases CLIP carries — a caveat the paper does not dwell on.