Beyond Accuracy: A Multi-dimensional Cognitive Audit of Medical Large Vision–Language Models in Fundus Image Interpretation

Reliance on standalone accuracy limits credible assessment of fundus-focused large vision–language models (LVLMs), as high scores often stem from linguistic shortcuts rather than real visual reasoning. This work develops the Cognitive Audit Framework (CAF), a four-module automated auditing pipeline that dissects model reasoning flaws: Visual–Linguistic Decoupling (textual dependency via modality ablation), Hierarchical Logical Consistency (lesion–diagnosis contradiction detection), Reasoning Fidelity Gap (chain-of-thought unfaithfulness scoring), and Contextual Robustness (positional bias under option permutation). Experiments on six 7B–31B LVLMs over FunBench reveal a notable gap between benchmark accuracy and reasoning quality: high accuracy coexists with measurable textual dependency, logical inconsistencies across diagnostic levels, limited chain-of-thought faithfulness, and non-trivial positional sensitivity. CAF serves as a reproducible complement to pure accuracy metrics for validating clinical competence of ophthalmic multimodal models.