VERICITE: Evaluating Sentence-Level Citation Faithfulness in Retrieval-Augmented Medical Question Answering

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Health & Medical Research · Depth: Expert, quick

Summary

VERICITE is a novel framework designed to evaluate sentence-level citation faithfulness in retrieval-augmented generation (RAG) systems, specifically within medical question answering. Addressing the challenge that RAG does not guarantee citations support associated claims, VERICITE operates by retrieving PubMed abstracts via the NCBI E-Utilities API, prompting large language models (LLMs) to generate answers with inline citations, and then verifying each citation using a DeBERTa-v3-large NLI model. The framework was applied to evaluate four LLMs on 500 BioASQ questions, testing retrieval depths of 3 and 5, with extended experiments up to k=15, and an oracle setting. Results indicate that only 27-41% of citation pairs are supported at the sentence level at k=3 and k=5, with rates decreasing at larger k. Crucially, even under oracle conditions, while answer quality improved, citation faithfulness did not substantially, suggesting that LLM generation behavior significantly contributes to unfaithful citations.

Key takeaway

For AI Scientists and Machine Learning Engineers developing RAG systems for medical question answering, these findings highlight a critical challenge: your current RAG implementations likely produce unfaithful citations. You should integrate robust, sentence-level citation verification mechanisms, such as NLI models, into your evaluation pipelines. Furthermore, prioritize research and development into fine-tuning LLM generation strategies to directly improve citation accuracy, as simply increasing retrieval depth or using better documents alone will not resolve this core issue.

Key insights

RAG systems in medical QA exhibit low sentence-level citation faithfulness, often due to LLM generation behavior.

Principles

• Citation faithfulness in RAG requires sentence-level verification.
• LLM generation behavior significantly impacts citation accuracy.
• Increased retrieval depth does not improve citation faithfulness.

Method

VERICITE retrieves PubMed abstracts, prompts LLMs for inline-cited answers, then verifies each citation at the sentence level using a DeBERTa-v3-large NLI model.

In practice

• Implement sentence-level NLI models for RAG citation checks.
• Focus on LLM fine-tuning to improve citation generation.
• Evaluate RAG systems beyond answer quality to include faithfulness.

Topics

• Retrieval-Augmented Generation
• Citation Faithfulness
• Medical Question Answering
• Large Language Models
• Natural Language Inference
• PubMed Abstracts
• BioASQ

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

Related on AIssential

• Benchmarking Retrieval Strategies for Biomedical Retrieval-Augmented Generation: A Controlled Empirical Study — Cross-Encoder Reranking significantly improves biomedical RAG performance by enhancing contextual precision.
• Retrieval-Augmented Generation for Clinical Question Answering in Portuguese Drug Leaflets: Benefits and Limitations — RAG improves factual recall in clinical Q&A but needs adaptive control for complex reasoning to avoid anchoring bias.
• Vector RAG vs LLM-Compiled Wiki: A Preregistered Comparison on a Small Multi-Domain Research — Grounded research synthesis involves multiple capabilities, with no single architecture excelling across all metrics.
• Structured Summaries for Retrieval-Augmented Generation in Portuguese-Language Consumer Complaints — Structured summaries significantly improve dense retrieval effectiveness in RAG systems for noisy, lengthy texts.
• RARE: Redundancy-Aware Retrieval Evaluation Framework for High-Similarity Corpora — Existing RAG benchmarks misrepresent real-world performance due to high document redundancy and similarity in enterprise corpora.
• When Cases Get Rare: A Retrieval Benchmark for Off-Guideline Clinical Question Answering — LLMs require retrieval augmentation and strong reasoning for reliable off-guideline medical question answering.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.