Researchers propose a new approach called 'future querying' to use large language models (LLMs) as implicit medical world models. This involves evaluating LLMs' ability to answer time-indexed clinical queries about a patient's future using unstructured clinical documentation. The framework uses endpoint-agnostic training, allowing a single model to answer diverse clinical queries without manual feature engineering or task-specific retraining. Experiments on synthetic and real-world data show promising results, with small fine-tuned models matching or approaching larger proprietary systems. --- Why it matters: This matters because it could enable the development of more efficient and privacy-preserving medical prediction models that don't require extensive manual feature engineering or large amounts of curated structured data. Source: https://arxiv.org/abs/2608.23248