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

2026-03-05 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, extended

Summary

MA-RAG (Multi-Round Agentic RAG) is introduced to enhance medical reasoning in Large Language Models (LLMs) by mitigating issues like hallucinations and outdated knowledge, which traditional Retrieval-Augmented Generation (RAG) methods struggle with due to noisy token-level signals and lack of multi-round refinement. This framework employs an agentic refinement loop that iteratively evolves both external evidence and internal reasoning history, transforming semantic conflict among candidate responses into actionable retrieval queries. MA-RAG comprises a Solver Agent for generating diverse candidates, a Retrieval Agent for conflict-guided evidence fetching, and a Ranking Agent to optimize history reasoning traces, thereby addressing "long-context degradation" and extending the "self-consistency principle" by using inconsistency as a proactive signal. Evaluated across seven diverse medical Q&A benchmarks, MA-RAG consistently outperformed competitive inference-time scaling and RAG baselines, achieving a substantial +6.8 points average accuracy improvement over the Llama-2-7B-Chat backbone model, reaching 69.7%. The framework demonstrated particular effectiveness on complex tasks like MedXpertQA, showing a 37% improvement over the backbone, with ablation studies confirming the critical contributions of its agentic components and multi-round refinement.

Key takeaway

MA-RAG, a Multi-Round Agentic RAG framework, significantly boosts LLM medical reasoning by iteratively refining responses and external evidence. It leverages semantic conflict among candidate answers to guide targeted retrieval and optimizes reasoning history, achieving a +6.8 average accuracy improvement over backbone models and a 37% gain on complex MedXpertQA. This approach enhances reliability by mitigating hallucinations and long-context degradation, crucial for safety-critical medical AI.

Topics

• Medical Reasoning
• Retrieval-Augmented Generation
• Agentic AI
• Multi-Round Refinement
• Test-Time Scaling

Code references

• NJU-RL/MA-RAG

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• SEMA-RAG: A Self-Evolving Multi-Agent Retrieval-Augmented Generation Framework for Medical Reasoning — Multi-agent RAG with self-evolving, sufficiency-driven retrieval significantly improves medical question answering accuracy.
• Quantifying the Accuracy and Cost Impact of Design Decisions in Budget-Constrained Agentic LLM Search — Budgeting search depth and retrieval quality offers more accuracy gains than expanding token limits for agentic RAG.
• TechGraphRAG: An Agentic Graph-Augmented RAG Framework for Technical Literature Reasoning — Agentic RAG frameworks can enhance technical reasoning by integrating multi-step pipelines, knowledge graphs, and external search for robust evidence grounding.
• MARDoc: A Memory-Aware Refinement Agent Framework for Multimodal Long Document QA — MARDoc uses a structured, dynamically updated memory to overcome context dilution in multimodal long-document QA agents.
• LLM-Based Multi-Agent System with Retrieval-Augmented Generation for Medical Care Planning Generation in Sickle Cell Disease — RAG-enhanced multi-agent LLMs can improve safety and clinical quality in high-risk medical care planning.
• MADRAG: Multi-Agent Debate with Retrieval-Augmented Generation for Training-Free Analytic Essay Scoring — Multi-agent debate with retrieval augmentation offers a training-free, reliable approach to analytic essay scoring, outperforming prompt-based LLMs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.