FACTORY: A Challenging Human-Verified Prompt Set for Long-Form Factuality

2025-07-31

AI Summary

Existing long-form factual evaluation benchmarks largely lack human verification, compromising reliability and hindering accurate assessment of language models’ ability to generate factually accurate and complete answers under short prompts. Method: We introduce FactLong—the first large-scale, human-verified benchmark for long-text factual evaluation—designed to stress-test tail-fact reasoning. FactLong employs model-assisted question construction followed by multi-round human refinement to ensure answerability, unambiguity, and truth-seeking; systematic evaluation is grounded in expert human annotation. Contribution/Results: Experiments reveal that state-of-the-art language models exhibit a 40% factual error rate on FactLong—four times higher than on conventional benchmarks (~10%)—demonstrating its significantly greater difficulty. FactLong thus establishes a more reliable, discriminative, and methodologically rigorous standard for evaluating factual consistency in long-context generation.

Technology Category

• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Long-form factuality evaluation assesses the ability of models to generate accurate, comprehensive responses to short prompts. Existing benchmarks often lack human verification, leading to potential quality issues. To address this limitation, we introduce FACTORY, a large-scale, human-verified prompt set. Developed using a model-in-the-loop approach and refined by humans, FACTORY includes challenging prompts that are fact-seeking, answerable, and unambiguous. We conduct human evaluations on 6 state-of-the-art language models using FACTORY and existing datasets. Our results show that FACTORY is a challenging benchmark: approximately 40% of the claims made in the responses of SOTA models are not factual, compared to only 10% for other datasets. Our analysis identifies the strengths of FACTORY over prior benchmarks, emphasizing its reliability and the necessity for models to reason across long-tailed facts.

Problem

• Evaluates long-form factuality in model responses
• Addresses lack of human-verified benchmarks
• Identifies high factual error rates in SOTA models

Innovation

• Human-verified large-scale prompt set
• Model-in-the-loop development approach
• Challenging fact-seeking unambiguous prompts