TLDR: A new study introduces HIVMedQA, a benchmark to evaluate large language models (LLMs) for HIV medical decision support. It tested ten LLMs on HIV-related questions, assessing comprehension, reasoning, knowledge recall, bias, and harm. Gemini 2.5 Pro consistently outperformed others. Key findings include that proprietary models often performed better, medically fine-tuned models didn’t always surpass general-purpose ones, and LLMs struggled more with reasoning and comprehension than factual recall. The study also found LLMs susceptible to cognitive biases and highlighted the effectiveness of LLM-as-a-judge evaluation over lexical matching for clinical accuracy.
Large language models (LLMs) are rapidly changing how we approach many tasks, and healthcare is no exception. These advanced AI systems are increasingly seen as valuable tools to help doctors with their daily decisions. Managing HIV, a complex and ever-evolving condition with diverse treatment options, other health issues, and adherence challenges, is a particularly compelling area where LLMs could offer significant support.
However, bringing LLMs into actual clinical practice comes with its own set of hurdles. Concerns about accuracy, the potential for harm, and whether clinicians will accept these tools are major considerations. Despite their promise, the performance of AI in HIV care hasn’t been thoroughly investigated, and there’s a clear lack of studies specifically benchmarking LLMs in this field.
A recent study aimed to address this gap by evaluating the current state of LLMs for HIV management, highlighting their strengths and limitations. To do this, the researchers developed a new benchmark called HIVMedQA. This benchmark is designed to assess how well LLMs answer open-ended medical questions related to HIV patient management. The dataset for HIVMedQA consists of carefully selected HIV-related questions, which were developed and validated with the help of an infectious disease physician.
The study put ten different LLMs to the test, including seven general-purpose models and three specialized medical LLMs. They used a technique called prompt engineering to optimize how the models responded. To measure performance, they used and expanded upon several scoring methods, including lexical similarity (how much the words in the answer match a reference) and an innovative approach called ‘LLM-as-a-judge.’ This ‘LLM-as-a-judge’ method involved using a powerful LLM (GPT-4o) to evaluate the answers of other models, extending existing metrics to better capture the subtle nuances important in medicine.
The evaluation focused on several key aspects: how well the models understood the questions (comprehension), their ability to reason through clinical scenarios, their recall of medical knowledge, any potential biases, the possibility of causing harm, and factual accuracy.
The findings revealed that Gemini 2.5 Pro consistently outperformed all other models across most of these dimensions. Interestingly, two of the three top-performing models were proprietary, meaning their inner workings are not publicly disclosed. The study also found that strong performance was limited to only a few LLMs, especially as the complexity of clinical questions increased. Surprisingly, medically specialized models didn’t always perform better than general-purpose ones, and simply having a larger model size (more parameters) wasn’t a reliable indicator of effectiveness.
The researchers also observed that tasks requiring reasoning and deep comprehension posed greater challenges for LLMs compared to simple knowledge recall. This suggests that while LLMs are good at remembering facts, they still struggle with complex problem-solving. Furthermore, the models were found to be susceptible to common cognitive biases, such as recency bias (giving more weight to recent information), frequency bias (believing something is more common because it’s encountered more often), and status quo bias (preferring the current state of affairs).
A significant insight from the study was that evaluating responses using an LLM-as-a-judge proved more effective at capturing clinical accuracy than traditional lexical matching methods. This is because medical answers can be correct even if they use different wording than a reference answer, and an LLM-as-a-judge can understand the semantic meaning better.
These findings underscore the critical need for more targeted model development and evaluation strategies to ensure that LLMs can be safely and effectively integrated into clinical decision support systems. For more detailed information, you can read the full research paper available at arXiv.
Also Read:
- Bridging the Health Data Divide: Why AI Medical Benchmarks Need an African Focus
- Assessing How Large Language Models Understand Structured Information
The study recommends several improvements for future LLM development in healthcare. Evaluation frameworks should move beyond simple factual recall to assess clinical reasoning and how models handle uncertainty. Instead of just lexical similarity, methods like LLM-as-a-judge are crucial for capturing true clinical accuracy. Current medical fine-tuning strategies, which often focus only on injecting static knowledge, are insufficient; future approaches should enhance reasoning, comprehension of complex cases, and robustness against biases. Finally, training datasets need to reflect the diverse and often ambiguous nature of real-world clinical cases to better prepare models for practical use in varied healthcare settings.
