Study finds episodic and hybrid context strategies lead on long-distance clinical reasoning

A paper submitted to arXiv on 30 Sep 2026 examines whether large language models can answer clinical questions using evidence scattered through lengthy patient histories. The authors frame this as a problem of finding and combining pertinent details distributed across an extended record. They caution that increasing the amount of history supplied to a model does not automatically make necessary information easier to use or improve its reasoning.

To test context design, the study compares Full, Recent, Episodic, Semantic and Hybrid strategies on MedLoCoMo, using four open-weight LLMs. The evaluation measures correctness, performance as the gap between a question and its supporting evidence widens, and abstention when a prompt contains an unsupported premise. This setup considers both whether a model answers supported questions and how it behaves when the record cannot support the requested conclusion.

Across the comparison, Episodic and Hybrid produced the strongest overall accuracy in general, according to the abstract. They also retained the top accuracy when supporting details were far from the questions, while Recent Context showed the steepest performance drop as that separation increased. The reported results make distance between question and evidence an explicit part of the comparison, rather than treating all context as equally accessible.

The authors also tested adversarial cases in which a question assumes a conclusion that the available patient history does not substantiate. They found that strong results on answerable questions did not necessarily coincide with successful abstention on these unsupported cases. The abstract therefore presents abstention as a distinct challenge alongside accuracy, rather than as something assured by strong answers to supported questions.

For AI builders, the findings make evidence selection and presentation a key design consideration in systems that reason across long histories, not merely the quantity of context they can accept. The paper's conclusion is that reliable longitudinal reasoning depends critically on how relevant evidence is chosen and shown to the model. The study's comparison gives builders a basis for examining context approaches against both distant supporting evidence and unsupported-premise questions, the two challenges described in its evaluation.