What Medical Large Language Models Remember: Benefits, Risks, and Recommendations

This study examined what large language models memorize when adapted for medicine, how frequently memorization occurs, how much content is retained, and how it may affect downstream medical applications. The researchers evaluated memorization across continued pretraining on medical corpora, fine-tuning on standard medical benchmarks, and fine-tuning on real-world clinical records.

Adapting large language models to medicine is intended to improve their medical knowledge and task performance, but the same training may also cause them to reproduce content from their training data. Some memorization may be useful, such as accurate recall of clinical guidelines, while other forms may add little value or expose sensitive patient information. Understanding these differences is essential for developing effective, generalizable, and responsible medical AI.

The researchers evaluated both general-purpose and medically adapted language models across ten datasets and three common adaptation settings: continued pretraining on medical corpora, fine-tuning on medical question-answering benchmarks, and fine-tuning on real-world clinical data. Memorization was measured using exact, approximate, semantic, and task-specific measures, supplemented by manual review of thousands of model outputs. The clinical case study included more than 13,000 unique inpatient records from Yale New Haven Health System.

Memorization was prevalent across all three adaptation settings and took different forms at different stages of training. The researchers identified three broad types: beneficial memorization, such as accurate recall of clinical guidelines and biomedical concepts; uninformative memorization, such as boilerplate or templated language; and harmful memorization, such as benchmark-specific content and sensitive clinical information.

Memorization was also persistent: up to 87% of content memorized during continued pretraining remained after subsequent fine-tuning. In the clinical case study, fine-tuning improved diagnostic performance but also increased the model’s tendency to reproduce training-record content, including 3,192 verified instances of protected health information across 10,000 evaluated outputs.

The findings show that memorization in medical AI is not uniformly beneficial or harmful. Retaining accurate medical knowledge may support model performance, while surface-level copying may limit generalization and the reproduction of sensitive clinical content may create privacy risks. Medical AI should therefore be evaluated not only by its accuracy, but also by what it retains and may reproduce from its training data.

The authors recommend evaluating memorization alongside conventional performance metrics throughout model development. Potential strategies include improving the quality of training data, reducing duplicate and repetitive content, using privacy-preserving training methods, applying appropriate access controls to models trained on clinical data, and strengthening reporting standards. They also call for further research on how to preserve useful medical knowledge while minimizing uninformative and harmful memorization.