Neuropsychology, Vol 40(7), Oct 2026, 697-711; doi:10.1037/neu0001079
Objective: We developed and evaluated a training-free, large language model (LLM) multiagent framework, consisting of role-prompted LLM instances, that simulates hospital-style team diagnosis for Alzheimer’s disease, improving interpretability and generalizability for multimodal data. Method: We designed a self-learning and critical multiagent workflow comprising nine single-dimension agents and five senior agents that iteratively validate, retrieve experience, and refine decisions via self-evaluation and critique. Agents are augmented with an “intelligence” module and an experience-retrieval module. Participants were 1,362 Alzheimer’s Disease Neuroimaging Initiative (ADNI) records spanning cognitively normal, mild cognitive impairment, and Alzheimer’s disease from ADNI GO, ADNI 1, ADNI 2, ADNI 3, ADNI 4. The decision layer is not supervised on ADNI labels, and no end-to-end fine-tuning is performed on ADNI within this study. Results: The full self-learning and critical multiagent framework achieved an F1 score = 88%, sensitivity = 88%, and precision = 90% for three-way cognitively normal/mild cognitive impairment/Alzheimer’s disease classification, providing case-level rationales. Relative to early agents, sequential self-learning and critique produced large effect gains. Single-dimension agents varied widely, while the integrated framework yielded the most balanced performance and higher answer relevancy on LLM-based metrics. Conclusions: A label-free LLM multiagent approach can integrate heterogeneous modalities, retain diagnostic “experience,” and generate transparent reasoning while delivering competitive accuracy on ADNI. The framework shows promise for real-world support of complex neuropsychological diagnosis; future work should incorporate fully multimodal LLMs for imaging and repeated, graph-driven iterations to further enhance performance. (PsycInfo Database Record (c) 2026 APA, all rights reserved)