Truth or Mirage? Towards End-to-End Factuality Evaluation with LLM-Oasis

TL;DR

This work introduces LLM-Oasis, the largest resource for end-to-end factuality evaluation built from Wikipedia by extracting atomic claims, subtly falsifying one, and generating paraphrased factual and unfactual texts. It enables two benchmarking tasks: end-to-end factuality evaluation and evidence-based claim verification, validated with a gold standard through extensive human annotation. A modular pipeline trained on LLM-Oasis—comprising claim extraction, evidence retrieval, and claim verification—achieves competitive performance against large LLMs, with a small, fine-tuned model reaching $69.24\%$ accuracy on the end-to-end task and a dedicated verifier achieving $93.30\%$ on evidence-based verification. The results underscore the difficulty of factuality evaluation, motivate architectural modularity over monolithic LLMs, and point to future work expanding domain and language coverage to enhance practical factuality assessment tools.

Abstract

After the introduction of Large Language Models (LLMs), there have been substantial improvements in the performance of Natural Language Generation (NLG) tasks, including Text Summarization and Machine Translation. However, LLMs still produce outputs containing hallucinations, that is, content not grounded in factual information. Therefore, developing methods to assess the factuality of LLMs has become urgent. Indeed, resources for factuality evaluation have recently emerged. Although challenging, these resources face one or more of the following limitations: (i) they are tailored to a specific task or domain; (ii) they are limited in size, thereby preventing the training of new factuality evaluators; (iii) they are designed for simpler verification tasks, such as claim verification. To address these issues, we introduce LLM-Oasis, to the best of our knowledge the largest resource for training end-to-end factuality evaluators. LLM-Oasis is constructed by extracting claims from Wikipedia, falsifying a subset of these claims, and generating pairs of factual and unfactual texts. We then rely on human annotators to both validate the quality of our dataset and to create a gold standard test set for benchmarking factuality evaluation systems. Our experiments demonstrate that LLM-Oasis presents a significant challenge for state-of-the-art LLMs, with GPT-4o achieving up to 60% accuracy in our proposed end-to-end factuality evaluation task, highlighting its potential to drive future research in the field.

Figures (3)

• Figure 1: Pipeline for the creation of LLM-Oasis. Given a passage from a Wikipedia page (original text on top), we task an LLM to: extract a list of atomic claims (1), falsify one of the extracted claims (2), and then, given the two sets of claims, produce a paraphrase of the original text (3), and an alternative version featuring the unfactual information (4).
• Figure 2: Recall@k performance of the $E5_{base}$ model at different values of $k$.
• Figure 3: Balanced Accuracy (%) of different models in the RAG setting with increasing context length. Each model is evaluated on inputs of 2k, 4k, 8k, 16k, and 18k tokens, which marks the length of the longest prompt instantiated in our evaluation.