Cell子刊：龙尔平/万沛星团队发布大模型“圆桌会议”框架，大幅提升医疗AI推理能力

Core Viewpoint - The article discusses the rapid development of medical AI, particularly focusing on the limitations of single models and the introduction of the "Model Confrontation and Collaboration" (MCC) framework to enhance medical reasoning and decision-making in AI systems [3][4]. Group 1: MCC Framework - The MCC framework aims to transition medical AI from "single-point intelligence" to "collaborative reasoning" by creating a dynamic, debate-based model that enhances reliability, interpretability, and collaboration [4]. - It establishes a "shared context workspace" where different language models can generate answers and key arguments in parallel, ensuring that all models have visibility into the complete dialogue history during debates [8]. Group 2: Core Process of MCC - The MCC process consists of three main steps: 1. Independent Reasoning: Multiple models generate answers and key arguments simultaneously, with a gate system activating debates only when discrepancies arise [9]. 2. Debate as Action: Models engage in multi-round messaging, performing actions like questioning, providing evidence, rebutting, and reflecting on their reasoning chains to enhance accuracy [10]. 3. Consensus Optimization: After each round, a consensus is determined, and if no agreement is reached within three rounds, a majority vote is used as the fallback output strategy [10]. Group 3: Performance Metrics - The MCC framework has demonstrated strong performance on various medical benchmarks, achieving an average accuracy of 92.6% on MedQA and maintaining over 90% accuracy across multiple subjects in MMLU [13]. - It also shows robust performance in more challenging assessments, such as achieving approximately 40% accuracy on MedXpertQA and effectively managing uncertainty in MetaMedQA [14]. Group 4: Long-form Question Answering - In long-form question-answering tasks, the MCC framework outperformed other models in key dimensions, with improvements of 8-12 percentage points in critical areas such as reasoning correctness and bias control [16]. - The framework achieved a comprehensive score of 92.1 on HealthBench, indicating its robustness and safety in complex clinical scenarios [16]. Group 5: Interactive Diagnostic Conversations - The MCC framework was tested in simulated diagnostic conversations, capturing over 80% of key patient information points and demonstrating higher relevance in questioning compared to single models [19]. - In diagnostic conclusions, the MCC framework achieved an 80% accuracy rate in preferred diagnoses, showcasing its ability to enhance diagnostic reasoning through collaborative questioning [19]. Group 6: Implications and Future Directions - The study indicates that multi-model confrontation and collaboration can enhance medical reasoning capabilities without additional task training or external knowledge bases, improving the quality and stability of outputs in complex scenarios [22]. - The MCC framework is not intended to replace physicians but to provide multi-faceted arguments and traceable debate logs to assist clinical personnel in reducing diagnostic errors and enhancing decision transparency [22].