arxivcs.CVcs.AI2026-07-10
TCLA: Training-Free Class-wise Logit Adaptation for Medical Vision-Language Models
Medical Vision-Language Models (VLMs) exhibit strong zero-shot performance, yet their effectiveness still declines on out-of-distribution (OOD) data due to domain shifts and class bias inherited from large-scale pretraining. Existing few-shot adaptation methods typically introduc…