Medical knowledge evolves continuously, while the parametric knowledge stored in large language models is fixed at training time. External retrieval such as retrieval‑augmented generation can bring in newly published evidence, but the retrieved items may be irrelevant, incomplete, or contradictory, which can hurt factual accuracy of LLM outputs.
To tackle this, we introduce CLEAR, an agentic framework for cross‑source evidence adjudication in medicine. The system first generates candidate answers independently through three complementary pathways: the model’s internal parametric knowledge, a locally curated corpus, and dynamically retrieved literature.
An aggregation verifier then jointly evaluates the candidates, their supporting evidence, provenance, and source‑quality signals, spotting agreement and conflict across sources.
Based on this, an adjudication module decides whether to keep the current conclusion or revise it, using an override‑guard and a challenge‑audit mechanism. Unresolved conflicts trigger targeted follow‑up search and re‑adjudication.
This design systematically reconciles multi‑source evidence, improving factual reliability and traceability of medical LLMs. Review