LLM-Cite: Cheap Fact Verification with Attribution via URL Generation
Keywords: Fact Verification, Attribution, Citation, Factuality
Abstract: Hallucinations are one of the main issues with Large Language Models (LLMs). This has led to increased interest in automated ways to verify the factuality of LLMs' responses. Existing methods either rely on: (a) search over a knowledge base (KB), which is costly especially if the KB must be updated frequently to keep up with fresh content, (b) LLM's parametric knowledge to fact-check claims, which is cheaper but does not give attribution and is limited to verifying claims related to knowledge acquired during pretraining. In this work, we present LLM-Cite, a cheap and easy to implement method that does not rely on any external search system while still providing attribution and the ability to verify fresh claims. Our key insight is to leverage an LLM to directly generate potential citation URLs for a given claim, and then use entailment checks to verify the claim against content of the URLs (which are fetched on-the-fly). We benchmark LLM-Cite on three datasets containing fresh and non-fresh claims generated by humans and models. We show that LLM-Cite performs comparable or better than existing methods on all categories of claims --- importantly, without sacrificing attribution, or requiring costly external search --- overall LLM-Cite is more than 45x cheaper than a Google Search based approach.
Primary Area: interpretability and explainable AI
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 12217
Loading