With the rapid advancement of multimodal foundation models, artificial intelligence (AI) has shown growing potential in computational pathology. However, current approaches still rely heavily on large annotated datasets, while real-world pathological data are limited and generalize poorly across cohorts. To address these issues, we systematically evaluate state-of-the-art large multimodal models (LMMs) on pathology-related tasks under zero-shot and few-shot settings. We further propose a retrieval-augmented support-set prompting framework that incorporates class-aware retrieved image-text exemplars to enhance contextual reasoning and pathological image understanding. By using class-aware retrieved image-text exemplars as external evidence, the framework enables models to capture fine-grained pathological patterns and produce more consistent and interpretable predictions. Extensive experiments on multiple pathology datasets demonstrate significant performance gains, highlighting the effectiveness of retrieval-based contextual augmentation in improving both reasoning ability in complex medical domains. The code and datasets are available at https://github.com/ttchu1221/RALM-Path .
Chu et al. (Tue,) studied this question.