TCLA: Training-Free Class-wise Logit Adaptation for Medical Vision-Language Models

Medical vision-language models perform well in zero-shot settings but degrade when applied to new imaging domains due to distribution shifts and class imbalance from pretraining. This paper introduces a training-free method that adjusts inference logits using only a handful of support examples, improving class separation without adding trainable components - important since low-data regimes (like one-shot) are often unstable for existing adaptation techniques. Tested across nine datasets spanning X-ray, ultrasound, MRI, CT, and histopathology, it outperforms most training-based alternatives. Applications include rapid, lightweight adaptation of medical AI diagnostic tools across imaging modalities and hospital settings with minimal labeled data. Authors: Tianyou Jiang, Ziyu Zhou Paper: https://arxiv.org/abs/2607.09562v1