Future Querying: Can LLMs Serve as Implicit Medical World Models?

TL;DR - This paper introduces “future querying,” which tests whether LLMs can act as implicit medical world models by answering time-indexed questions about patient trajectories from unstructured clinical notes. It could support flexible, privacy-preserving clinical prediction without separate pipelines for every endpoint.

• Uses endpoint-agnostic training so one model can answer diverse future-oriented clinical queries without manual feature engineering or task-specific retraining.
• Evaluates the approach on synthetic medical reports and real ICU notes from MIMIC-IV.
• Finds that small, locally fine-tuned open-weight models can match or approach larger proprietary systems.
• Results suggest LLMs can capture aspects of clinical dynamics while remaining suitable for on-premise deployment.