Abstract
Medical vision-language models (VLMs) offer promise for clinical decision support, yet their reliability under distribution shifts remains a major concern for safe deployment. These models often learn task-agnostic correlations due to variability in imaging protocols and free-text reports, limiting their generalizability and increasing the risk of failure in real-world settings. We propose DRiFt, a structured feature decoupling framework that explicitly separates clinically relevant signals from task-agnostic noise using parameter-efficient tuning (LoRA) and learnable prompt tokens. To enhance cross-modal alignment and reduce uncertainty, we curate high-quality, clinically grounded image-text pairs by generating captions for a diverse medical dataset. Our approach improves in-distribution performance by +11.4% Top-1 accuracy and +3.3% Macro-F1 over prior prompt-based methods, while maintaining strong robustness across unseen datasets. Ablation studies reveal that disentangling task-rel