Researchers have developed a framework called MACD for large language models (LLMs) to self-learn clinical knowledge. The Multi-Agent Clinical Diagnosis (MACD) framework allows LLMs to summarize, refine, and apply diagnostic insights in a multi-agent pipeline that mirrors the professional development of human physicians. This approach is tested on 4,390 real-world patient cases across seven diseases, showing an average improvement of 11.6 percentage points over established authoritative knowledge. The study also explores a collaborative workflow between LLM-based agents and human oversight, achieving an 18.3-percentage-point improvement over physician-only diagnosis. --- Why it matters: This work matters to AI researchers because it presents a scalable self-learning paradigm that bridges the gap between LLMs' intrinsic knowledge and real-world clinical practice demands. The proposed framework has significant implications for developing reliable, interpretable, and deployable AI-assisted diagnosis systems. Source: https://arxiv.org/abs/2509.20067