VeReaFine: Iterative Verification Reasoning Refinement RAG for Hallucination-Resistant on Open-Ended Clinical QA

Pakawat Phasook, Rapepong Pitijaroonpong, Jiramet Kinchagawat, Amrest Chinkamol, Tossaporn Saengja, Kiartnarin Udomlapsakul, Jitkapat Sawatphol, Piyalitt IttichaiwongACL 2025

TL;DR

This paper introduces VeReaFine, a novel "Verifier-RAG" pipeline that iteratively retrieves, verifies, and refines information to combat hallucinations in open-ended medical question answering. By integrating a dedicated verifier at multiple stages, VeReaFine enables even moderately sized language models to produce more factually accurate and evidence-grounded responses, rivaling larger models.

What Problem Does This Paper Address?

Large Language Models (LLMs) are increasingly being explored for medical applications, but a significant challenge is their tendency to "hallucinate" – generating plausible-sounding but incorrect clinical assertions. This poses a serious risk to patient safety and trust. While Retrieval-Augmented Generation (RAG) helps by providing external context, it doesn't guarantee the LLM will correctly use or adhere to this information. Existing methods like Chain-of-Thought prompting or self-verification often struggle with incomplete knowledge or lack systematic integration of external evidence, especially in complex domains like medicine where factual accuracy is paramount. The paper aims to bridge this gap by creating a system that rigorously verifies every claim against retrieved evidence.

Methods

VeReaFine employs a multi-stage, iterative pipeline that can be summarized as a "tri-loop" of retrieve, verify, and generate/refine.

  1. Initial Retrieval: The process begins by encoding the user's medical query and retrieving the top-k relevant passages from a large biomedical corpus (PubMed, StatPearls) using a two-stage dense retriever and reranker.
  2. Context Verification Loop: A small, specialized medical reasoning verifier (8B parameters) assesses the direct relevance of retrieved passages. It identifies "ground-truth" context and discards irrelevant snippets. Crucially, if the current context is insufficient, the verifier formulates a targeted "feedback query" to fetch missing evidence. This loop can run up to three times.
  3. Answer Generation: A 7B-parameter generator (Qwen2.5-7B-Instruct) is then conditioned solely on the verified "ground-truth" context to draft an initial answer.
  4. Hallucination Check & Refinement: The verifier re-examines the generated draft, flagging any unsupported claims. If hallucinations are detected, the generator is prompted to revise or remove them. This iterative refinement ensures that every assertion is evidence-backed.

This design emphasizes explicit verification at each step, adaptive retrieval based on identified information gaps, and a constraint-satisfaction approach to ensure factuality.

Datasets

The experiments were conducted on the BioNLP 2025 ClinIQLink test set. This dataset is specifically designed to evaluate hallucinations in open-ended medical QA and includes four formats: short answer, short-inverse, multi-hop, and multi-hop-inverse. The dataset provides ground truth and requires evidence-grounded answers, making it suitable for assessing hallucination resistance.

Key Insights and Ideas

  • Verifier as a Gatekeeper: The core innovation is the explicit integration of a verifier throughout the RAG process, acting as a critical gatekeeper to filter out irrelevant information and flag unsupported claims.
  • Adaptive Retrieval: The ability of the verifier to generate targeted "feedback queries" for missing evidence is key to efficiently expanding context only when necessary, avoiding the pitfalls of overwhelming the generator with irrelevant data.
  • Small Models, Big Impact: VeReaFine demonstrates that even a moderately sized LLM (7B parameters) can achieve performance comparable to or exceeding much larger models (32B) when augmented with a robust verification mechanism. This highlights the potential for efficient, factually reliable AI assistants.
  • Iterative Refinement for Factuality: The iterative nature of the verification and refinement loop, akin to an Expectation-Maximization process, is designed to monotonically improve factuality by systematically addressing unsupported claims.

Key Results

  • On the BioNLP 2025 ClinIQLink challenge, VeReaFine (7B generator) achieved recall gains of +60–100% at the 75th percentile (P75) compared to the same model without verification.
  • It recovered over 90% of the step-identification fidelity of a 32B baseline model on multi-hop inverse questions.
  • In short-inverse questions, VeReaFine achieved a P75 of 0.559, significantly outperforming both the base 7B model (0.358) and a 32B model (0.539).
  • For multi-hop inverse questions, VeReaFine showed a +22% boost over the 32B model in inverse step performance.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Benchmark Against Standard RAG: A direct comparison against a standard RAG pipeline (without the verifier loop) on the same corpus is needed to quantify the verifier's marginal benefit.
  • Scaling to Larger Models: Investigating VeReaFine's performance with larger LLMs (30B-70B) could reveal whether verification benefits plateau or diminish at scale.
  • Multi-Verifier Ensembles: Exploring ensembles of diverse verifier models (e.g., NLI models, external fact-checkers) could enhance robustness and capture a wider range of hallucination types.
  • Human-in-the-Loop: Incorporating clinician feedback could help calibrate verifier thresholds and ensure clinical relevance.
  • Efficient Verification: Developing lightweight verifier architectures or using knowledge distillation could reduce latency for real-time clinical applications.

Summary updated

BibTeX
@inproceedings{phasook2025vereafine,
  author = {Pakawat Phasook and Rapepong Pitijaroonpong and Jiramet Kinchagawat and Amrest Chinkamol and Tossaporn Saengja and Kiartnarin Udomlapsakul and Jitkapat Sawatphol and Piyalitt Ittichaiwong},
  title = {{VeReaFine: Iterative Verification Reasoning Refinement RAG for Hallucination-Resistant on Open-Ended Clinical QA}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure