Large language models (LLMs) are increasingly explored for clinical decision support, and medical multiple‑choice questions (MCQs) serve as a natural testbed. Options often share overlapping evidence, with only a few decisive factors separating the correct answer. Existing retrieval‑based or reasoning‑oriented approaches mainly supply broadly relevant knowledge or generate reasoning traces, but they do not isolate the compact differential evidence that distinguishes confusing alternatives. This work introduces a retrieval‑grounded differential evidence framework. For each option it extracts option‑specific evidence fragments and assembles them into concise evidence prompts used for answer selection. The design aligns external knowledge with differential reasoning while cutting irrelevant context. Experiments across several datasets and target LLMs show consistent accuracy gains. Ablation studies reveal that the framework outperforms directly using retrieved evidence chunks or closed‑source LLM‑generated general hints, and that the improvements are not solely due to answer‑proximal guidance but stem from differential evidence acting as a compact aid for reasoning over competing choices. The findings also offer insights for medical reasoning and other multi‑option decision‑making settings.
Review: Extracting and packaging differential evidence bridges retrieval and reasoning effectively, and appears to be a key factor in boosting LLM performance on medical MCQs.