Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Benchmarking LLM Faithfulness in RAG with Evolving Leaderboards

Manveer Singh Tamber, Forrest Sheng Bao, Chenyu Xu, Ge Luo, Suleman Kazi, Minseok Bae, Miaoran Li, Ofer Mendelevitch, Renyi Qu, Jimmy Lin

TL;DR

This work tackles the persistent problem of LLM hallucinations in retrieval-augmented generation by proposing two complementary efforts: a live hallucination leaderboard based on the HHEM detector and an LLM-as-a-judge framework called FaithJudge that leverages human-annotated hallucination examples to improve automated evaluation. FaithJudge demonstrates higher agreement with human judgments than existing detectors and extends beyond summarization to other RAG tasks via the RAGTruth dataset, broadening the scope of faithfulness benchmarking. The authors provide a live FaithJudge repo, compare various detectors, and show that larger LLM judges and more annotated examples enhance detection, though challenges remain, especially for benign and questionable labels. Overall, the paper offers a scalable, domain-adaptive approach to benchmarking and improving LLM faithfulness in RAG, with practical implications for building more trustworthy generative AI systems.

Abstract

Retrieval-augmented generation (RAG) aims to reduce hallucinations by grounding responses in external context, yet large language models (LLMs) still frequently introduce unsupported information or contradictions even when provided with relevant context. This paper presents two complementary efforts at Vectara to measure and benchmark LLM faithfulness in RAG. First, we describe our original hallucination leaderboard, which has tracked hallucination rates for LLMs since 2023 using our HHEM hallucination detection model. Motivated by limitations observed in current hallucination detection methods, we introduce FaithJudge, an LLM-as-a-judge framework that leverages a pool of diverse human-annotated hallucination examples to substantially improve the automated hallucination evaluation of LLMs. We introduce an enhanced hallucination leaderboard centered on FaithJudge that benchmarks LLMs on RAG faithfulness in summarization, question-answering, and data-to-text generation tasks. FaithJudge enables a more reliable benchmarking of LLM hallucinations in RAG and supports the development of more trustworthy generative AI systems: https://github.com/vectara/FaithJudge.

Benchmarking LLM Faithfulness in RAG with Evolving Leaderboards

TL;DR

This work tackles the persistent problem of LLM hallucinations in retrieval-augmented generation by proposing two complementary efforts: a live hallucination leaderboard based on the HHEM detector and an LLM-as-a-judge framework called FaithJudge that leverages human-annotated hallucination examples to improve automated evaluation. FaithJudge demonstrates higher agreement with human judgments than existing detectors and extends beyond summarization to other RAG tasks via the RAGTruth dataset, broadening the scope of faithfulness benchmarking. The authors provide a live FaithJudge repo, compare various detectors, and show that larger LLM judges and more annotated examples enhance detection, though challenges remain, especially for benign and questionable labels. Overall, the paper offers a scalable, domain-adaptive approach to benchmarking and improving LLM faithfulness in RAG, with practical implications for building more trustworthy generative AI systems.

Abstract

Retrieval-augmented generation (RAG) aims to reduce hallucinations by grounding responses in external context, yet large language models (LLMs) still frequently introduce unsupported information or contradictions even when provided with relevant context. This paper presents two complementary efforts at Vectara to measure and benchmark LLM faithfulness in RAG. First, we describe our original hallucination leaderboard, which has tracked hallucination rates for LLMs since 2023 using our HHEM hallucination detection model. Motivated by limitations observed in current hallucination detection methods, we introduce FaithJudge, an LLM-as-a-judge framework that leverages a pool of diverse human-annotated hallucination examples to substantially improve the automated hallucination evaluation of LLMs. We introduce an enhanced hallucination leaderboard centered on FaithJudge that benchmarks LLMs on RAG faithfulness in summarization, question-answering, and data-to-text generation tasks. FaithJudge enables a more reliable benchmarking of LLM hallucinations in RAG and supports the development of more trustworthy generative AI systems: https://github.com/vectara/FaithJudge.
Paper Structure (14 sections, 3 figures, 7 tables)

This paper contains 14 sections, 3 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Vectara's Original Hallucination Leaderboard
  4. FaithBench
  5. FaithJudge
  6. Evaluating Hallucination Detectors
  7. Evaluation Datasets
  8. Existing Hallucination Detectors
  9. Evaluating FaithJudge
  10. Adding More Evaluation Tasks
  11. Conclusion
  12. Summaries with FaithJudge Verdicts
  13. Judge Bias
  14. FaithJudge Rankings

Figures (3)

  • Figure 1: Proportion of summary FaithBench labels (left) and FaithJudge predictions (right) across models. For FaithBench labels, red indicates Unwanted, orange indicates Questionable, yellow indicates Benign, while green indicates Consistent. For FaithJudge predictions, red indicates Hallucinated, and green indicates Consistent summaries. Each bar shows the proportion of summaries falling into each category.
  • Figure 2: Sensitivity and specificity with FaithJudge as the number of examples in the prompt is increased. We place an asterisk (*) next to the 10 because, in this case, FaithJudge is shown annotations for the summary it is evaluating.
  • Figure 3: Two LLM summaries of a news source from FaithBench with FaithJudge hallucination judgements. The red text indicates portions of the summaries that contain hallucinations.