Despite the demonstrated promise of Large Language Models in medical question answering, existing work largely addresses closed-form, exam-style tasks and overlooks complex open-ended questions requiring reasoning over noisy, long clinical documents. In this work, we present our system, OptiMed, submitted to the ArchEHR-QA 2026 shared task on grounded clinical question answering over EHR notes. We combine GEPA, an evolutionary prompt optimization framework, with multi-agent majority voting across five diverse LLMs and a structured clinical abstraction strategy for question interpretation. OptiMed ranked 1st overall among teams completing all four subtasks with an average score of 52.0, achieving top AlignScore in both Question Interpretation and Answer Generation, reflecting strong factual grounding. GEPA optimization proved effective for structured tasks with sufficient development data, but failed to generalize on complex generative tasks under very limited number of supervisions. Multi-agent majority voting consistently lifted performance in evidence-oriented subtasks. Prompt analysis attributes GEPA’s gains to role prompting and procedural decomposition and failures to over-specification under limited supervision.

Despite the impressive reported performance of Large Language Models (LLMs) on medical Question-Answering (QA), evaluation still overrelies on multiple-choice and extractive settings primarily assessing textbook knowledge, with limited work on open-ended QA over patient records. We present a comprehensive study investigating LLM long-context (LC) comprehension on longitudinal, patient-centric Electronic Health Records (EHRs) across multiple datasets, models and metrics, covering extractive, multiple-choice, and open-ended generative QA under different context formulation and Retrieval Augmented Generation (RAG) settings. We show that medically fine-tuned models offer no measurable benefit over base counterparts, smaller models exhibit stronger memorization, and longitudinal supplementary notes aid generative QA. Reasoning models gain on complex multi-note QA under LLM Correctness but not stricter reference-aligned metrics, while hybrid RAG with reranking outperforms Full-Context (FC) prompting. A question-category analysis and multi-faceted metric evaluation further reveal where RAG and FC excel and struggle, with RAG falling behind in longitudinal reasoning, thus uncovering systematic failures pointing to the need for reasoning-appropriate metrics.