Large language models (LLMs) have demonstrated impressive performance across various domains. However, for clinical diagnosis, higher expectations are required for LLM’s reliability and sensitivity: thinking like physicians and remaining sensitive to key medical information that affects diagnostic reasoning, as subtle variations can lead to different diagnosis results. Yet, existing works focus mainly on investigating the sensitivity of LLMs to irrelevant context and overlook the importance of key information. In this paper, we investigate the sensitivity of LLMs, i.e. GPT-3.5, GPT-4, Gemini, Claude3 and LLaMA2-7b, to key medical information by introducing different perturbation strategies. The evaluation results highlight the limitations of current LLMs in remaining sensitive to key medical information for diagnostic decision-making. The evolution of LLMs must focus on improving their reliability, enhancing their ability to be sensitive to key information, and effectively utilizing this information. These improvements will enhance human trust in LLMs and facilitate their practical application in real-world scenarios. Our code and dataset are available at https://github.com/chenwei23333/DiagnosisQA.

In the field of dialogue summarization, due to the lack of training data, it is often difficult for supervised summary generation methods to learn vital information from dialogue context with limited data. Several attempts on unsupervised summarization for text by leveraging semantic information solely or auto-encoder strategy (i.e., sentence compression), it however cannot be adapted to the dialogue scene due to the limited words in utterances and huge gap between the dialogue and its summary. In this study, we propose a novel unsupervised strategy to address this challenge, which roots from the hypothetical foundation that a superior summary approximates a replacement of the original dialogue, and they are roughly equivalent for auxiliary (self-supervised) tasks, e.g., dialogue generation. The proposed strategy RepSum is applied to generate both extractive and abstractive summary with the guidance of the followed nˆth utterance generation and classification tasks. Extensive experiments on various datasets demonstrate the superiority of the proposed model compared with the state-of-the-art methods.