CRITIC-RAG: Knowledge-Augmented Large Language Models With Verified Retrieval for Improved Medical Reasoning

While retrieval-augmented generation (RAG) presents a promising solution for enhancing large language models (LLMs) in question answering (QA), particularly in knowledge-intensive domains like medicine, it continues to face challenges related to factual consistency and reasoning reliability. To address this challenge, we propose CRITIC-RAG, a verification-enhanced framework that integrates a small-size, instruction-tuned verifier throughout the RAG pipeline. Our method incorporates selective retrieval, evidence filtering, structured reasoning through self-consistency, and groundedness verification. This approach enhances the accurate utilization of external knowledge while reducing spurious generations. Comprehensive experiments across five medical QA benchmarks and multiple LLM backbones confirm the framework's broad applicability and plug-and-play adaptability. For instance, CRITIC-RAG improves accuracy from 41.3% to 48.8% on MedQA and boosts BERTScore from 68.1% to 74.4% on MedicationQA using LLaMA-3. Ablation studies further reveal that evidence filtering and structured reasoning are especially critical to robust performance. Case studies and retrieval analyses, including improvements in entailment-based similarity scores and factual accuracy, demonstrate how each stage of verification jointly contributes to generating more accurate, evidence-grounded responses. Overall, our work highlights verification as a key mechanism for enhancing trustworthiness in knowledge-intensive medical QA, a capability that is increasingly important for clinical decision support and other real-world healthcare applications.