Large Language Models (LLMs) are increasingly being used for medical advice by patients and healthcare providers. These models capture knowledge from their training data, which consists of vast medical corpora. However, they lack the ability to use this knowledge to causally reason about the underlying physiological processes. Moreover, they are unable to deal with uncertainty, generating responses that are confidently presented yet factually incorrect. Acting on such factually incorrect medical advice can be dangerous. Mitigating these risks requires rethinking the role of LLMs in medicine. In this work, we present an evaluation scheme for LLMs in three roles: direct clinical decision support, exact medical knowledge base, and approximate medical knowledge base. We evaluate six LLMs on two clinical studies, in obstetrics and pediatric critical care, respectively. Our results indicate that LLMs are much better suited to the approximate knowledge base role. Based on these observations, we request caution when directly employing LLMs in safety-critical domains such as medicine.
We study model merging as a practical alternative to standard adaptation for code-mixed NLP. Starting from a multilingual base model, we (i) perform continued pre-training (CPT) on unlabeled code-mixed text to obtain an adapted checkpoint, (ii) merge this checkpoint with the base model, and (iii) fine-tune (FT) on downstream task data. We evaluate this approach on sentence classification (sentiment and hate speech) in English-Hindi (En-Hi) and English-Spanish (En-Es) using XLM-R and Llama-3.2-1B. Merged models consistently outperform both full fine-tuning and CPT.FT, yielding gains of 2-5 F-1 points over full fine-tuning and similar to 1-2 over CPT.FT, suggesting that merging exploits unlabeled data more effectively than CPT alone. Zero-/few-shot prompting with larger LLMs (e.g., Llama-3.3-70B) trails fine-tuned and merged checkpoints, highlighting limits of in-context learning for codemixed inputs. Cross-pair transfer from En-Hi to En-Ta/Ml shows merged checkpoints (e.g., TV/TIES) outperform monolingual baselines (0.65-0.68 vs. 0.61-0.63 F-1), confirming code-mixed knowledge as a superior substrate for low-resource pairs. We conclude with adaptation recipes for common data regimes (labeled only; labeled+unlabeled; transfer-only) and discuss limitations and scaling considerations for broader tasks and larger models.
