CiteGuard: Faithful Citation Attribution for LLMs via Retrieval-Augmented Validation
Yee Man Choi, Xuehang Guo, Yi R., Fung, Qingyun Wang
2025-10-24
Summary
This paper investigates how well large language models (LLMs) do at correctly citing sources when they write scientific text, and proposes a new method to improve this accuracy.
What's the problem?
LLMs are becoming helpful for writing, but a big concern is whether the citations they include are actually correct and support the information presented. Simply asking another LLM to check the citations, which is a common approach, isn't very reliable either, because the checking LLM can also make mistakes. The core issue is making sure the citations an LLM uses are the same ones a human researcher would use for the same content.
What's the solution?
The researchers came up with a system called CiteGuard. Instead of just having an LLM judge the citations, CiteGuard actively *searches* for relevant sources to back up the LLM's writing. It's like giving the LLM a research assistant to double-check its work. This 'retrieval-aware' approach means CiteGuard doesn't just rely on what it already 'knows,' but actively looks for evidence to support the citations. This improved accuracy by over 12% and reached a level comparable to human performance on a standard test called CiteME.
Why it matters?
This work is important because it addresses a critical flaw in LLMs – their tendency to make up or misattribute sources. By improving citation accuracy, we can trust LLMs more when they're used for scientific writing and research, and CiteGuard can even suggest alternative, valid citations that the LLM might have missed.
Abstract
Large Language Models (LLMs) have emerged as promising assistants for scientific writing. However, there have been concerns regarding the quality and reliability of the generated text, one of which is the citation accuracy and faithfulness. While most recent work relies on methods such as LLM-as-a-Judge, the reliability of LLM-as-a-Judge alone is also in doubt. In this work, we reframe citation evaluation as a problem of citation attribution alignment, which is assessing whether LLM-generated citations match those a human author would include for the same text. We propose CiteGuard, a retrieval-aware agent framework designed to provide more faithful grounding for citation validation. CiteGuard improves the prior baseline by 12.3%, and achieves up to 65.4% accuracy on the CiteME benchmark, on par with human-level performance (69.7%). It also enables the identification of alternative but valid citations.