Measuring and Reducing LLM Hallucination Without Gold-Standard Answers

LLM hallucination, i.e. generating factually incorrect yet seeminglyconvincing answers, is currently a major threat to the trustworthiness andreliability of LLMs. The first step towards solving this complicated problem isto measure it. However, existing hallucination metrics require having abenchmark dataset with gold-standard answers, i.e. "best" or "correct" answerswritten by humans. Such requirements make hallucination measurement costly andprone to human errors. In this work, we propose Factualness Evaluations viaWeighting LLMs (FEWL), an innovative hallucination metric that is specificallydesigned for the scenario when gold-standard answers are absent. FEWL leveragesthe answers from off-the-shelf LLMs that serve as a proxy of gold-standardanswers. The key challenge is how to quantify the expertise of reference LLMsresourcefully. We show FEWL has certain theoretical guarantees and demonstrateempirically it gives more accurate hallucination measures than naively usingreference LLMs. We also show how to leverage FEWL to reduce hallucinationthrough both in-context learning and supervised fine-tuning. Extensiveexperiment results on Truthful-QA, CHALE, and HaluEval datasets demonstrate theeffectiveness of FEWL.