HiTZ-IXA at ArchEHR-QA 2026: Evidence Alignment Through Self-Consistency and Prompt Curation in Memory-Constrained Environments

Abstract

The development of question-answering systems capable of grounding their answers in Electronic Health Records could provide patients with faithful assistance while reducing the clinical workload. The ArchEHR-QA 2026 Shared Task was organized to advance progress in this context. In this paper, we present our strategies for addressing this shared task, which are focused primarily on evidence alignment and, to a lesser extent, on evidence identification. Our approaches rely exclusively on open-source models with up to 8 billion parameters, aiming to produce systems suitable for environments with memory constraints. We experimented with methods based on embedding models, prompt curation, self-consistency, and combination of LLMs. We concluded that prompt curation together with an effective post-processing step was crucial for creating stable systems, while self-consistency yielded considerable gains in performance. The results of our approaches suggest that small LLMs can substantially improve their accuracy in the evidence alignment task via simple and affordable techniques.

Resources

Details