HallucinationHunter: Fine-Grained Factual Grounding of Generated Text

Peter Belcak, Yu Zhang, Shizhe Diao, David Eric Austin, Yonggan Fu, Ryan Angilly, Yingyan Celine Lin, Eileen Margaret Peters Long, Pavlo Molchanov

Keywords: reliability, hallucinations, hallucination detection, fact checking, human factors in NLP, factuality

TL;DR: An applied compound AI system for fine-grained claim attribution, hallucination detection, and removal, including a demo, library, and a REST server

Abstract: Language models often ignore or directly contradict their inputs. We present HallucinationHunter (HH), a first-principles, model-agnostic factual attribution and hallucination elimination system. HH systematically attributes all claims made in model outputs to segments in provided ground truth sources and surgically marks or removes unsupported claims. This offers significantly more grounding information and fine control in factual generation than the standard guardrail accept/reject sample assessment. Across Google FACTS, FActScore, RAGTruth, and FalseNeedles benchmarks, we show that HH consistently improves the factuality of model responses in the groundedness setting without resorting to any sample rejections. We make HH available both as a library and a REST endpoint server for easy integration in factuality-sensitive agentic applications. Our implementation of HH is described in detail in the appendix together with an interactive demo, which we plan to publish when de-anonymized.

Submission Number: 62