Researchers at Yale School of Medicine have developed a new benchmarking framework to evaluate emerging artificial intelligence models designed to support complex clinical decision-making. Called MedicalAgentsBench, the framework compares two rapidly evolving approaches to medical reasoning by large language models (LLMs): systems in which multiple AI models collaborate like a multidisciplinary medical team, and advanced single models that perform complex reasoning internally before producing an answer.
The researchers believe the benchmark could help healthcare organizations assess future clinical decision support systems more effectively, particularly as hospitals increasingly develop their own AI models to meet privacy and regulatory requirements.
AI-assisted clinical reasoning
Clinical decision-making often involves specialists from different disciplines combining their expertise to diagnose patients or determine the most appropriate treatment. Inspired by this multidisciplinary approach, researchers have begun developing AI systems that simulate collaborative medical discussions.
One example is MedAgents, introduced by the Yale team in 2023. The system assigns different large language models the roles of medical specialists, such as surgeons, radiologists or pathologists. These AI agents exchange information and debate possible answers over several rounds before reaching a consensus.At the same time, another generation of AI models has emerged that performs similar multi-step reasoning within a single model. These so-called internal reasoning models use reinforcement learning to work through complex problems step by step before generating a response.
According to the researchers, both approaches aim to improve the quality of medical reasoning, but until now there has been no standardized way to compare their strengths and limitations.
A benchmark beyond memorization
Existing evaluation methods often rely on questions like those found in medical licensing or admissions examinations. However, many modern AI models now achieve such high scores on these tests that it has become increasingly difficult to distinguish genuine reasoning ability from simple memorization.
To address this limitation, the Yale researchers developed MedicalAgentsBench, a benchmarking framework containing more than 800 challenging clinical questions derived from eight established medical datasets. Rather than testing factual recall, the questions require multiple reasoning steps to arrive at a correct answer. The researchers designed the benchmark to identify how effectively different AI architectures reason through unfamiliar clinical scenarios, rather than how well they reproduce information encountered during training.
Using MedicalAgentsBench, the team compared collaborative multi-agent systems with internally reasoning models. The results did not identify a clear overall winner. Instead, both approaches demonstrated complementary strengths. The study also found that combining the two methods, allowing multiple external AI agents to work alongside an internally reasoning model, could further improve overall performance.
Future clinical AI development
The researchers suggest MedicalAgentsBench could become a practical evaluation tool for healthcare organizations developing their own clinical AI systems. Because patient privacy regulations often prevent hospitals from using publicly available large language models for clinical decision support, many institutions are expected to build or fine-tune models using their own secure infrastructure.
In that context, benchmarking frameworks that assess not only performance but also computational cost and implementation complexity may help organizations make informed decisions when selecting AI architectures. The authors emphasize that their work is not intended to determine which AI model should be used in clinical practice. Instead, MedicalAgentsBench provides a standardized method for evaluating different reasoning strategies as medical AI continues to evolve.
By shifting the focus from traditional knowledge testing to complex clinical reasoning, the framework reflects the growing ambition of AI developers to create systems capable of supporting healthcare professionals in increasingly sophisticated diagnostic and treatment decisions. The researchers expect the benchmark to contribute to the development of more robust, transparent and clinically useful AI-based decision support tools in the future.
Generative AI and LLMs in healthcare
Several generative AI tools and large language model (LLM) based solutions are already deployed within healthcare. Although the added value of such solutions has been described and demonstrated several times, the quality of some tools still falls short of the high demands in healthcare. This was also recently shown in an Israeli study. It concluded that most large language models (LLMs), such as ChatGPT, still underperform for medical decision-making.
Despite the growing popularity of AI in healthcare, the research shows that these models often provide inaccurate or inconsistent information, posing risks when making medical decisions. The researchers stress that although AI has potential, human expertise remains essential in the diagnostic process. They therefore recommend using AI tools only as a support tool for now, and not as a replacement for medical professionals.
References
Publication:journal Patterns
