TLDR: FACTORY is a new, large-scale, human-verified dataset designed to rigorously test the long-form factual accuracy of large language models (LLMs). Unlike previous benchmarks, FACTORY uses a model-in-the-loop approach combined with extensive human annotation to create challenging, fact-seeking prompts. Evaluations show that state-of-the-art LLMs perform significantly worse on FACTORY compared to other datasets, highlighting its effectiveness in identifying current limitations in LLM factuality and reasoning capabilities.
In the rapidly evolving world of artificial intelligence, Large Language Models (LLMs) are becoming increasingly sophisticated, capable of generating extensive and coherent text. However, a persistent challenge remains: ensuring the factual accuracy of these long-form generations. While much research has focused on improving LLM factuality, the tools used to evaluate this crucial aspect often fall short, primarily due to a lack of human oversight in their creation.
Addressing this critical gap, researchers from FAIR at Meta have introduced a groundbreaking new benchmark called FACTORY. This large-scale, human-verified prompt set is specifically designed to rigorously test the ability of LLMs to produce accurate and comprehensive responses to prompts that require detailed, long-form answers. Unlike many existing benchmarks that rely on automated generation, FACTORY incorporates a unique ‘model-in-the-loop’ approach combined with extensive human verification to ensure its quality and challenging nature.
The creation of FACTORY began with seed topics drawn from Wikipedia, which were then expanded into a diverse array of prompts using LLMs. To maintain a high level of difficulty, a filtering process was applied to remove simpler prompts. The most crucial step involved human annotators who meticulously refined each prompt. These human experts ensured that every prompt was fact-seeking, answerable using publicly available information, unambiguous, not time-sensitive, and safe. This rigorous human touch is what sets FACTORY apart, making its evaluation results significantly more trustworthy.
The FACTORY dataset includes a ‘hard’ split, comprising 421 prompts, which presents a substantial challenge to even the most advanced LLMs. When state-of-the-art models like Claude 3.7 Sonnet, Gemini 2.5 Pro, and GPT-4o were benchmarked against FACTORY, the results were striking. While these models achieved around 90% factual precision on existing benchmarks like LongFact and FactBench, their performance on FACTORY’s hard subset dropped significantly to approximately 60%. This stark difference underscores FACTORY’s effectiveness in identifying the true limitations of current LLMs in generating factually precise long-form content.
The researchers also delved into why FACTORY proves to be so challenging. They found that previous benchmarks often suffered from issues such as unanswerable questions, subjective prompts, time-sensitive information, or even prompts leading to hallucinations. In contrast, FACTORY’s prompts are notably longer and demand more specific details, requiring LLMs to reason across a wide array of ‘long-tailed facts’ – less common or more obscure pieces of information. Experiments with ‘atomic prompts’ (breaking down complex prompts into simpler, single-fact questions) further revealed that LLMs struggle not only with the complexity of long prompts but also with the underlying knowledge required to answer them.
Also Read:
- R1-ACT: A New Method for Activating Safety Knowledge in AI Reasoning Models
- RL-PLUS: A New Approach to Expand LLM Reasoning Capabilities Beyond Current Limits
In essence, FACTORY sets a new standard for evaluating long-form factuality in LLMs. By emphasizing human verification and requiring models to synthesize detailed, specific information, it paves the way for future advancements in model development and evaluation, pushing the boundaries of what AI can factually achieve. For more details, you can refer to the research paper.
