MedAgentBench v2: Improving Medical LLM Agent Design

Abstract

MedAgentBench is the first benchmark for evaluating LLM agents on clinical tasks in a FHIR-compliant EHR. In this paper, we present significant prompt engineering and tool design improvements over the original agent implementation and introduce a memory component that enables the agent to learn from prior failures. We added new tools for the agent to properly format its output for tasks, interact with an EHR without constructing explicit HTTP requests, which were prone to syntax errors, and make math calculations. We also wrote a new system prompt that asked the agent to outline its plan before making any tool calls and think step by step using chain of thought reasoning, and provided few shot examples of good vs. bad outputs. Using GPT-4.1 as the base model, our agent achieved a success rate of 91.0% without memory and 98.0% with memory. A surprising consequence is that the agent performed better on a different task that had no associated memory entry, possibly demonstrating that LLMs can adapt to the style of tasks presented by users. To contribute to the benchmark and evaluate the generalization of our agent, we developed 300 new multi-step clinically-driven tasks in collaboration with a physician. Lastly, we show the current limitations of these benchmarks and highlight the necessary next steps and challenges for the responsible deployment of AI agents in real-world healthcare settings. We hope that this paper leads to further development of EHR agents and benchmarks.

Ähnliche Arbeiten