Supplementary materials for: Text-dominant decision-making by large multimodal models in dermatology clinical challenges

Supplemental Table 1: Metadata of JDCR case challenge
Supplemental Table 2.1: Prompts used to query LMMs
Supplemental Table 2.2: Answers from ChatGPT-o1 (“o1-2024-12-17”) to Case Challenge under three input conditions.
Supplemental Table 2.3: Answers from ChatGPT-5.2 (“gpt-5.2-2025-12-11”) to Case Challenge under three input conditions.
Supplemental Table 2.4: Answers from Gemini 3 (“gemini-3-pro-preview”) to Case Challenge under three input conditions.
Supplemental Table 2.5: Answers from Mistral (“pixtral-large-2411”) to Case Challenge under three input conditions.
Other Supplemental Results
+Supplemental Figure 1: Distribution of model-generated explanations in semantic embedding space
+Supplemental Table 3: Default API settings of each compared model
+Supplemental Table 4: Consistency of answer choice selection across runs
+Supplemental Table 5: Comparison of single-run and mean accuracies
+Supplemental Table 6: Analysis of image-anchored model explanations
+Appendix: Annotations of image-anchored explanations for 10 vignettes

补充表1：JDCR病例挑战赛元数据
补充表2.1：用于查询多模态大语言模型 (Large Multimodal Models, LMMs) 的提示词
补充表2.2：ChatGPT-o1（“o1-2024-12-17”）在三种输入条件下针对病例挑战赛的作答结果
补充表2.3：ChatGPT-5.2（“gpt-5.2-2025-12-11”）在三种输入条件下针对病例挑战赛的作答结果
补充表2.4：Gemini 3（“gemini-3-pro-preview”）在三种输入条件下针对病例挑战赛的作答结果
补充表2.5：Mistral（“pixtral-large-2411”）在三种输入条件下针对病例挑战赛的作答结果
其他补充结果
+ 补充图1：模型生成解释的语义嵌入空间分布
+ 补充表3：各对比模型的默认API设置
+ 补充表4：多次运行中答案选项选择的一致性
+ 补充表5：单次运行与平均准确率的对比分析
+ 补充表6：基于图像锚定的模型解释分析
+ 附录：10个示例片段的图像锚定解释标注