From Conflict to Consensus: Boosting Medical Reasoning via Multi-Round Agentic RAG

A team of researchers has introduced MA-RAG (Multi-Round Agentic RAG), a novel framework designed to significantly enhance the reliability of Large Language Models (LLMs) in high-stakes medical question-answering. The core innovation addresses a critical weakness: while standard Retrieval-Augmented Generation (RAG) mitigates hallucinations and outdated knowledge, it often relies on noisy, single-round retrieval. MA-RAG reframes the problem, treating a lack of consensus or 'conflict' among an LLM's own candidate answers as a proactive signal to initiate a multi-round, agent-driven refinement process. This moves beyond simple fact-checking towards a more dynamic, reasoning-focused loop.

The MA-RAG agent operates by first generating multiple candidate responses to a medical query. It then analyzes these for semantic conflict, using any disagreement to formulate precise, actionable queries for retrieving new external evidence. Simultaneously, it optimizes the internal reasoning history to combat performance degradation in long contexts. This iterative process, which the authors liken to a 'boosting' mechanism, minimizes residual error until a stable, high-fidelity consensus is reached. Extensive testing on 7 medical benchmarks shows MA-RAG consistently outperforms other inference-time scaling and RAG methods, delivering an average accuracy improvement of +6.8 points over the backbone model. This represents a meaningful step toward deployable, trustworthy AI assistants for clinical decision support by making reasoning transparent and evidence-driven.