Language Learning Models (LLMs) that are capable of interpreting natural language instructions to complete tasks are an exciting area of artificial intelligence research with direct implications for healthcare. Still, theypresent challenges as well. Researchers from Northeastern University and Codametrix conducted a study to evaluate the sensitivity of various LLMs to different natural language instructions specifically within the healthcare field. Their results revealed substantial variability in performance across all tested models, and confirmed that phrasing variations in instructions not only affect the outcome of tasks but also bring up issues of fairness in their responses.
Different types of LLMs have been enhanced in various ways to complete tasks with minimal or no previous examples or instructions. Some models, such as GPT-3.5+, FLAN, Alpaca, and Mistral, are designed for “zero-shot” task execution, meaning they can complete tasks they’ve never seen before simply by understanding natural language instructions. Despite these advances, LLMs tend to be highly sensitive to how these instructions are phrased. This sensitivity can be particularly challenging in specialized domains like healthcare, as minor variations in instructions can lead to major variations in the model’s performance.
In healthcare, these variations in performance can have significant consequences for patient treatment and care. For example, the study revealed that performance discrepancies exist between different demographic groups when conducting tasks like mortality prediction. These fairness considerations bring about a whole new area of concern when evaluating the applications of LLMs in clinical settings.
This research involved collecting prompts from medical doctors across various tasks and testing the sensitivity of seven general and specialized LLMs to the different phrasings of instructions. The researchers found that the models trained specifically on healthcare data tended to be more brittle and more sensitive to instruction variations than the general models.
Additionally, the research showed that even minor variations in phrasing could disproportionately impact demographic groups, further emphasizing the need for robustness and fairness in clinical LLMs. In general, the study concluded that general-domain LLMs outperformed domain-specific ones, even in healthcare-related tasks.
The performance of LLMs in healthcare is crucial due to the industry’s complexity and the potential impact on patient care. While domain-specific models were more brittle, general models performed measurably better across all tasks. Prompt variations impacting fairness are particularly concerning, given that the way instructions are phrased could lead to disparities in predictive accuracy among different demographic groups. In conclusion, the study highlights the urgent need for further research and development in this area to drive improvements in instruction-tuned LLM robustness.
