[Paper Review] Measuring and Reducing LLM Hallucination without Gold-Standard Answers

Jiaheng Wei, Yuanshun Yao+4 more|arXiv (Cornell University)|Feb 16, 2024

Plant-based Medicinal Research8 citations

TL;DR

FEWL is a hallucination metric designed for settings without gold-standard answers, using expert-weighted reference LLMs and laziness penalties, plus methods to reduce hallucination via ICL and LF-SFT, and a CHALE benchmark.

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.

Motivation & Objective

• Motivate measurement of LLM hallucination when gold-standard answers are unavailable.
• Introduce FEWL (F actualness Evaluations via Weighting LLMs) as an expert-weighted, proxy-based truthfulness metric.
• Provide theoretical guarantees linking FEWL to optimal answers under missing gold standards.
• Demonstrate how FEWL can be used to reduce hallucination through in-context learning and supervised fine-tuning.
• Release a large-scale CHALE benchmark to facilitate hallucination research under no-gold-standard settings.

Proposed method

• Compute an expertise-weighted truthfulness score by weighing reference LLM answers using lambda_i(x) derived from how reference LLMs disagree with intentionally wrong (IW) answers and agree with corrected (CO) answers.
• Introduce a laziness penalty by assessing how similar a reference LLM’s answers are across nearby questions within the same topic, penalizing superficially plausible but incorrect responses.
• Aggregate the two components using a variational form of f-divergence (e.g., Total-Variation) to produce FEWL.
• Algorithmic steps include querying multiple reference LLMs, generating IW/CO answers, computing lambda_i(x), computing laziness penalties on KNN questions, and combining via f-divergence form.
• Provide theoretical guarantees that FEWL ranks the best LLM highest, even without gold-standard answers.

Experimental results

Research questions

• RQ1 Can FEWL reliably measure hallucination without gold-standard answers by leveraging reference LLMs?
• RQ2 Do expert-weighted references and laziness penalties improve the accuracy of hallucination measurement compared to naïve reference-based metrics?
• RQ3 Can FEWL be used to reduce hallucination via in-context learning and supervised fine-tuning without gold-standard labels?
• RQ4 Does FEWL correlate with true non-hallucination rates and correctly rank LLMs by their hallucination propensity?
• RQ5 What is the impact of using CHALE as a large-scale benchmark for evaluating hallucination metrics without gold-standard data?

Key findings

• FEWL outperforms baselines in distinguishing non-hallucinated from hallucinated or half-hallucinated answers on CHALE and Truthful-QA.
• FEWL with expertise weighting and laziness penalty yields higher scores for non-hallucinated answers than baselines across GPT-3.5 and GPT-4 as reference LLMs.
• FEWL accurately ranks LLMs by their hallucination tendency, aligning with ground-truth non-hallucination rates on Truthful-QA.
• FEWL-guided in-context learning improves answer quality more than baseline ICL.
• Label-free supervised fine-tuning using FEWL-chosen answers improves win rates over pretrained baselines and approaches ground-truth labeling.