Source-linked AI summary
MedAgents: Large Language Models as Collaborators for Zero-shot Medical Reasoning
Xiangru Tang, Anni Zou, Zhuosheng Zhang, Ziming Li, Yilun Zhao, Xingyao Zhang, Arman Cohan, Mark Gerstein
TL;DR
Medical LLMs face limited specialized data and difficulty eliciting the domain knowledge and reasoning required for medical tasks. MedAgents uses role-playing agents and iterative multidisciplinary discussion to reach consensus without training or external knowledge bases. Across nine datasets, it surpasses zero-shot baselines and matches few-shot baselines, while human evaluation identifies error categories for future refinement.
Problem
Medical LLMs have limited specialized training data and struggle to elicit the extensive domain knowledge and reasoning needed for medical tasks.
Method
MedAgents uses role-playing LLM experts in five stages, including analysis, report synthesis, iterative consultation, and consensus decision making.
Results
Across nine datasets, MedAgents surpasses zero-shot baselines and matches few-shot baselines without prior training or dependence on knowledge bases.
Takeaways & Limitations
Role-playing and collaborative discussion provide a way to harness latent medical knowledge while supporting greater reliability and clarity in medical question answering.
Takeaways & Limitations
The framework may require ongoing knowledge updates, has limited applicability in low-resource languages, and costs more than chain-of-thought or direct prompting.
Abstract
from arXiv · showhide
Large language models (LLMs), despite their remarkable progress across various general domains, encounter significant barriers in medicine and healthcare. This field faces unique challenges such as domain-specific terminologies and reasoning over specialized knowledge. To address these issues, we propose MedAgents, a novel multi-disciplinary collaboration framework for the medical domain. MedAgents leverages LLM-based agents in a role-playing setting that participate in a collaborative multi-round discussion, thereby enhancing LLM proficiency and reasoning capabilities. This training-free framework encompasses five critical steps: gathering domain experts, proposing individual analyses, summarising these analyses into a report, iterating over discussions until a consensus is reached, and ultimately making a decision. Our work focuses on the zero-shot setting, which is applicable in real-world scenarios. Experimental results on nine datasets (MedQA, MedMCQA, PubMedQA, and six subtasks from MMLU) establish that our proposed MedAgents framework excels at mining and harnessing the medical expertise within LLMs, as well as extending its reasoning abilities. Our code can be found at https://github.com/gersteinlab/MedAgents.
1 Introduction
Medical question answering requires specialized domain knowledge and complex reasoning, while simple prompting can elicit hallucinations. MedAgents addresses this with multidisciplinary, role-playing collaboration intended to surface latent medical expertise and improve reasoning in a training-free, zero-shot setting.
- Motivation: Medical tasks demand domain expertise and advanced reasoning, while limited specialized training data and simple prompting hinder reliable medical reasoning.The paper identifies limited medical training data and difficulty eliciting medical expertise as major challenges.
- Motivation: MedAgents uses multidisciplinary LLM agents to expose expertise implicitly embedded in models and make it more accessible than traditional prompting.The framework draws on multi-agent collaboration and role-playing to harness latent knowledge.
- Framework: The framework reasons through expert gathering, analysis proposition, report summarization, collaborative consultation, and final decision making.During consultation, experts iteratively revise the report until approval or a maximum discussion threshold is reached.
- Evaluation: Experiments cover nine datasets, including MedQA, MedMCQA, PubMedQA, and six medical MMLU subtasks, primarily under zero-shot evaluation.The study compares the framework with chain-of-thought, self-consistency, and few-shot baselines.
- Error analysis: Human evaluation identifies four common error categories: missing domain knowledge, mis-retrieval, consistency errors, and chain-of-thought errors.The paper proposes targeted refinement of these shortcomings as a direction for improving proficiency and reliability.
2 Method
MEDAGENTS coordinates diverse medical agents through staged analysis, report synthesis, iterative consultation, and final decision-making. The framework uses question and option experts to produce and refine a consensus report.
- MEDAGENTS framework: MEDAGENTS operates through expert gathering, analysis proposition, report summarization, collaborative consultation, and final decision-making.The framework assembles domain experts, generates analyses, synthesizes them, revises the report through discussion, and derives a final answer.
- Expert Gathering: Question and option experts are recruited separately to provide diverse, multifaceted knowledge for each clinical question.Question-domain experts address the question, while option-domain experts analyze the answer choices in relation to the question and prior analyses.
- Analysis Proposition: Question analyses and option analyses are generated after expert recruitment, with option analysis incorporating both the choices and their relationship to the question.Question experts analyze the clinical question; option experts use the question, options, and question analyses for subsequent reasoning.
- Report Summarization: The report summarization stage synthesizes question and option analyses into a medical report containing extracted knowledge and an overall analysis.LLMs act as medical report assistants when converting the collected analyses into a synthesized report.
- Collaborative Consultation: Collaborative consultation repeatedly collects expert yes/no votes and modification opinions, revising the report until experts recognize it as consistent or the attempt limit is reached.Experts propose modifications when voting no, and the report is updated from those comments during iterative rounds.
3 Experiments
MEDAGENTS is evaluated zero-shot across nine medical datasets and compared with accessible prompting and sampling baselines. It outperforms zero-shot baselines by a large margin, while multi-agent role-playing mitigates performance degradation and hallucinations associated with isolated chain-of-thought prompting.
- Tasks and Datasets: The evaluation covers MedQA, MedMCQA, PubMedQA, and six medical subtasks from MMLU.The MMLU subtasks are anatomy, clinical knowledge, college medicine, medical genetics, professional medicine, and college biology.
- Implementation: All experiments use GPT-3.5-Turbo and GPT-4 in the zero-shot setting, with five question experts and two option experts except on PubMedQA.PubMedQA uses four question experts and two option experts; the framework uses five discussion iterations.
- Baselines: The comparisons include direct zero-shot and few-shot prompting, chain-of-thought prompting, and self-consistency sampling baselines.Self-consistency samples multiple chains from zero-shot or few-shot chain-of-thought and selects the majority answer.
- Limitations: The authors acknowledge that MEDAGENTS costs more than chain-of-thought or direct prompting, although they describe it as relatively cost-effective.The authors connect the improved performance's potential value to better health outcomes.
- Main Results: MEDAGENTS outperforms zero-shot baseline methods by a large margin across the evaluated datasets and achieves comparable zero-shot performance to few-shot approaches.Table 2 reports the main results across the nine datasets, comparing the framework with several baselines in zero-shot and few-shot settings.
- Main Results: Isolated chain-of-thought prompting can degrade performance through hallucinated medical terminology, whereas multi-agent role-playing mitigates these issues.The passage characterizes the hallucinations as spurious outputs associated with misapplied medical terminology.
4 Analysis
The analysis evaluates MEDAGENTS through ablations, agent-number and domain-variation studies, comparisons with open-source models, and human error analysis. It finds that role-playing and collaborative refinement improve performance, while domain knowledge remains the main error bottleneck.
- Baseline Comparisons: MEDAGENTS outperforms the baseline in Table 2 and comparable open-source methods fall short of that baseline.The open-source comparison motivates focusing on more effective methods.
- Agent-Number Studies: Introducing any number of expert agents significantly improves MedQA performance over the baseline.The study varies question and option agent counts while fixing other variables to examine performance trends.
- Error Analysis: 77% of 40 sampled MedQA and MedMCQA error cases involve domain-knowledge confusion, including missing or mis-retrieved knowledge.The four categories are lack of domain knowledge, mis-retrieval, consistency errors, and CoT errors.
- Correctional Capabilities and Interpretability: Collaborative consultation can correct earlier inaccuracies by synthesizing diverse expert opinions into more accurate conclusions.The paper presents this iterative refinement as evidence of correctional capability, interpretability, and accuracy in complex medical reasoning.
5 Related Work
Related work covers LLM applications in medicine, tool-augmented and instruction-tuned approaches, and the development of multi-agent collaboration. MEDAGENTS builds on role-playing and communication while emphasizing intrinsic medical knowledge without external knowledge bases.
- Medical LLM Applications: Medical LLM research spans applications including diagnostics, genetics, pharmacy, and medical evidence.These applications illustrate the growing use of LLMs in the medical field.
- Tool-Augmented Methods: Tool-augmented approaches acquire clinical information through external resources such as NCBI APIs, medical-guideline retrieval, or domain-specific fine-tuning.Examples include GeneGPT, Almanac, and KARD.
- Instruction Tuning: Instruction-tuned medical models commonly use external clinical knowledge bases and self-prompted data to construct fine-tuning datasets.Some target broad medical proficiency, while others focus on specialized areas or diverse medical instruction data.
- MEDAGENTS: MEDAGENTS differs from these approaches by harnessing latent medical knowledge through role-playing agents rather than relying on knowledge bases.Its design is inspired by multidisciplinary consultation in hospitals.
- LLM-based Multi-agent Collaboration: LLM-based multi-agent collaboration extends single-agent systems through multi-turn feedback, cooperation, role-playing, and communication.Prior work also explores adversarial collaboration through debate and negotiation.
6 Conclusion
The paper presents MEDAGENTS as a training-free medical QA framework using role-playing agents in multi-round discussions. Across nine datasets, it surpasses zero-shot baselines and matches few-shot baselines, while error evaluation identifies areas for refinement.
- Contribution: MEDAGENTS uses role-playing agents for multi-round discussions, providing greater reliability and clarity without prior training.The framework integrates medical knowledge through role-playing rather than depending on knowledge bases.
- Results: Across nine datasets, MEDAGENTS surpasses zero-shot baselines and matches few-shot baselines.The evaluated datasets include MedQA, MedMCQA, PubMedQA, and six MMLU subtasks.
Limitation
The framework has several unresolved boundaries that motivate future work, including knowledge updates, model integration, and low-resource language applicability.
- LLM parameterized knowledge may require updating over time to keep MEDAGENTS current.
- Future studies could investigate integrating diverse models at different stages of the framework.
- MEDAGENTS may have limited applicability in low-resource languages, motivating adaptation to a wider range of languages.
Ethics Statement
Although the work follows established medical question-answering benchmarks, applying LLM reasoning in medicine may introduce risks such as inherent model biases.
- The evaluation adheres to well-established medical question-answering benchmarks.
- Applying LLM reasoning to critical medical areas may introduce potential risks.
- Inherent biases in LLMs are identified as one example of these risks.
A Analysis on the Addition of CoT
The analysis reports that adding chain-of-thought can reduce zero-shot performance in specialist medical tasks, where hallucination and limited domain knowledge constrain its usefulness.
- Zero-shot CoT led to a performance drop compared with the zero-shot setting.
- In specialist domains requiring substantial expertise, CoT may sometimes produce hallucinated information.
- Prior medical question-answering work found that CoT’s step-by-step approach does not reliably generate correct medical answers.
- The study attributes most observed errors to missing domain knowledge rather than reasoning rationale, with 77% linked to domain knowledge and 8% to CoT.
- MEDAGENTS uses role-playing to help models reason with accurate knowledge without relying on retrieval-augmented generation.
B Interpretability: Case Study
The case studies illustrate how multidisciplinary analyses generate broader candidate interpretations and collaborative revision can move initial answers toward consensus and correction.
- Case-study mechanism: Different medical professionals contribute depth and breadth that transform initial analyses into comprehensive assessments.
- Eustachian tube case: Experts initially proposed tympanometry and the Politzer test for assessing Eustachian tube function.
- Eustachian tube case: Tympanometry was reaffirmed as a method for assessing Eustachian tube function, while additional methods included sonotubometry and nasal endoscopy.
- Eustachian tube case: The final report identified tympanometry as the best method for the patient’s condition after collaborative consultation.
- Embryology case: In the embryology case, experts proposed both amnion and chorion before discussion narrowed the conclusion toward amnion.
- Embryology case: The revised synthesis described amnion as the primary source of amniotic fluid while recognizing contributions from other structures.
C Dataset Information
The evaluation uses MedQA, MedMCQA, and medical-domain MMLU sub-subjects, with prompt templates and role descriptions documented separately. Repeated runs on 300 samples assess result variability and indicate stable reported performance.
- Dataset scope: MedQA contains USMLE-style questions with four or five answer choices, while MedMCQA contains four-option questions from Indian medical entrance examinations.
- Dataset scope: The study limits MMLU to sub-subjects within the medical domain, although MMLU broadly spans 57 subjects and educational levels.
- Experimental materials: Prompt templates and role descriptions used in the MedAgents framework are presented in Table 12.
- Repeated-run evaluation: Multiple runs on 300 samples were conducted for GPT-4 and GPT-3.5 to account for variability.Preliminary tests averaged scores across five repetitions per sample and found small variance between runs.
- Repeated-run evaluation: Table 13 compares GPT-4 and GPT-3.5 based on multiple runs of 300 samples.