Collaborative intelligence in AI: Evaluating the performance of a council of AIs on the USMLE

Abstract The variability in responses generated by Large Language Models (LLMs) like OpenAI’s GPT-4 poses challenges in ensuring consistent accuracy on medical knowledge assessments, such as the United States Medical Licensing Exam (USMLE). This study introduces a novel multi-agent framework—referred to as a "Council of AIs"—to enhance LLM performance through collaborative decision-making. The Council consists of multiple GPT-4 instances that iteratively discuss and reach consensus on answers facilitated by a designated "Facilitator AI." This methodology was applied to 325 USMLE questions across Step 1, Step 2 Clinical Knowledge (CK), and Step 3 exams. The Council achieved consensus responses that were correct 97%, 93%, and 94% of the time for Step 1, Step 2CK, and Step 3, respectively, outperforming single-instance GPT-4 models. In cases where there wasn’t an initial unanimous response, the Council of AI deliberations achieved a consensus that was the correct answer 83% of the time. For questions that required deliberation, the Council corrected over half (53%) of responses that majority vote had gotten incorrect. At the end of deliberation, the Council often corrected majority responses that were initially incorrect: the odds of a majority voting response changing from incorrect to correct were 5 (95% CI: 1.1, 22.8) times higher than the odds of changing from correct to incorrect after discussion. We additionally characterized the semantic entropy of the response space for each question and found that deliberations impact entropy of the response space and steadily decrease it, consistently reaching an entropy of zero in all instances. This study showed that in a Council model response variability—often viewed as a limitation—could be leveraged as a strength, enabling adaptive reasoning and collaborative refinement of answers. These findings suggest new paradigms for AI implementation and reveal diversity of responses as a strength in collective decision-making even in medical question scenarios where there is a single correct response. Author Summary In our study, we explored how collaboration among multiple artificial intelligence (AI) systems could improve accuracy on medical licensing exams. While individual AI models like GPT-4 often produce varying answers to the same question—a challenge known as "response variability"—we designed a "Council of AIs" to turn this variability into a strength. The Council consists of several AI models working together, discussing their answers through an iterative process until they reach consensus. When tested on 325 medical exam questions, the Council achieved 97%, 93%, and 94% accuracy on the Step 1, Step 2CK, and Step 3, respectively. This improvement was most notable when answers required debate: in cases where initial responses disagreed, the collaborative process corrected errors 83% of the time. Our findings suggest that collective decision-making— even among AIs—can enhance accuracy and AI collaboration can potentially lead to more trustworthy tools for healthcare, where accuracy is critical. By demonstrating that diverse AI perspectives can refine answers, we challenge the notion that consistency alone defines a "good" AI. Instead, embracing variability through teamwork might unlock new possibilities for AI in medicine and beyond. This approach could inspire future systems where AIs and humans collaborate (e.g. on Councils with both humans and AIs), combining strengths to solve complex problems. While technical challenges remain, our work highlights a promising path toward more robust, adaptable AI solutions.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.