Measuring and Reducing LLM Hallucination without Gold-Standard Answers

TL;DR

This work tackles measuring LLM hallucination in the absence of gold-standard answers by introducing FEWL, a metric that leverages multiple reference LLMs weighted by per-question expertise and a laziness penalty to generate a continuous hallucination score. FEWL constructs its score via a variational $f$-divergence framework, estimating per-question expertise $\lambda_i(x)$ from intentionally wrong/corrected answers and penalizing superficiality through proximity to neighboring questions. The authors provide theoretical guarantees showing FEWL can consistently favor the best-performing model in expectation, and they validate the approach with experiments on CHALE, Truthful-QA, and HaluEval, demonstrating accurate measurement and robust model/sample-level rankings. They further demonstrate practical utility by using FEWL to guide in-context learning and label-free supervised fine-tuning, achieving hallucination reduction at a fraction of the cost of human annotation, thus offering a scalable and cost-effective tool for trustworthy LLM deployment.

Abstract

LLM hallucination, i.e. generating factually incorrect yet seemingly convincing answers, is currently a major threat to the trustworthiness and reliability of LLMs. The first step towards solving this complicated problem is to measure it. However, existing hallucination metrics require having a benchmark dataset with gold-standard answers, i.e. "best" or "correct" answers written by humans. Such requirements make hallucination measurement costly and prone to human errors. In this work, we propose Factualness Evaluations via Weighting LLMs (FEWL), an innovative hallucination metric that is specifically designed for the scenario when gold-standard answers are absent. FEWL leverages the answers from off-the-shelf LLMs that serve as a proxy of gold-standard answers. The key challenge is how to quantify the expertise of reference LLMs resourcefully. We show FEWL has certain theoretical guarantees and demonstrate empirically it gives more accurate hallucination measures than naively using reference LLMs. We also show how to leverage FEWL to reduce hallucination through both in-context learning and supervised fine-tuning. Extensive experiment results on Truthful-QA, CHALE, and HaluEval datasets demonstrate the effectiveness of FEWL.

Figures (4)

• Figure 1: An overview of how to compute the FEWL (Factualness Evaluations via Weighting LLMs) score on an answer $y$ to a question $x$ when its golden-standard answer $y^*$ does not exist.
• Figure 2: In the Truthful-QA dataset, given a question $x$, and its top-10 most similar questions $x_1',\ldots x_{10}'$ with corresponding gold standard answers $y_1', \ldots, y_{10}'$, we show the fraction of times that these answers are judged (via GPT-4) to be a correct answer to the original question $x$.
• Figure 3: Human annotation on the three reference LLM answers: 0 indicates the wrong answer, 1 means partially correct, and 2 is the correct answer. Check 1, 2, and 3 denote Flan-Alpaca, GPT 3.5, and GPT 4, respectively.
• Figure 4: Boxplot of wrong and non-wrong score for 3 example LLMs. 'llm-name'-wrong-score: if the LLM believes $n$ out of 25 IW answers are correct, then $n$ will be the wrong score for this sample (larger $\rightarrow$ worse); 'llm-name'-non-wrong-score: if the LLM believes $n$ out of 25 CO answers are correct, then $n$ will be the non-wrong score for this sample (larger $\rightarrow$ better) (compare LLM answer with each IW/CO answer)

Theorems & Definitions (5)

• Theorem 3.4
• Remark 3.5
• Proposition B.1
• Remark B.2
• proof