FACTSCORE: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation

Sewon Min; Kalpesh Krishna; Xinxi Lyu; Mike Lewis; Wen-tau Yih; Pang Wei Koh; Mohit Iyyer; Luke Zettelmoyer; Hannaneh Hajishirzi

In brief

Evaluating whether long-form text generated by LLMs is factually accurate is challenging because generations contain many claims, some supported and some unsupported. FACTSCORE addresses this by breaking a generation into atomic facts (short statements with one piece of information each) and validating each against a knowledge source like Wikipedia. Human evaluation reveals that even state-of-the-art LLMs struggle with factuality: ChatGPT achieves only 58.3% and GPT-4 only 42.5%, far below human performance of 88.2%. The authors also introduce an automated FACTSCORE estimator based on retrieval and language models that approximates human judgments.

Contributions

Introduces FACTSCORE, a fine-grained evaluation metric that breaks long-form generations into atomic facts and validates each against a knowledge source, enabling per-fact precision assessment
Conducts extensive human evaluation of factual precision on 6,500 model-generated biographies, revealing systematic errors in commercial LLMs (InstructGPT 42.5%, ChatGPT 58.3%, Perplexity.AI 71.5% FACTSCORE)
Proposes an automated FACTSCORE estimator combining retrieval, language model scoring, and nonparametric methods, achieving <2% error rate compared to human judgments and enabling evaluation at scale without human effort
Evaluates 13 recent LLMs and demonstrates that larger models like Alpaca and Vicuna approach ChatGPT-level factuality, while smaller models remain significantly less factual

Method

FACTSCORE operates in two stages. First, a generation is decomposed into atomic facts — short sentences each containing one piece of verifiable information (typically 2-5 sentences per biography). This decomposition can be done manually or, for the automated estimator, by using InstructGPT with specific prompts. Second, each atomic fact is validated against Wikipedia by checking whether it is supported by the knowledge source. The validation yields a binary label: Supported if Wikipedia confirms the fact, Not-supported otherwise.

The automated estimator breaks facts into atomic units and validates them through four variants: (1) No-context LM — uses an LM to directly predict True/False; (2) Retrieve→LM — retrieves Wikipedia passages and has the LM validate based on retrieved context; (3) Nonparametric Probability (NP) — uses masked LM likelihood scores; and (4) Retrieve→LM + NP — ensemble of retrieval and NP approaches. The system uses LLAMA 7B and GPT3-powered retrieval to obtain knowledge source passages.

Results

Human evaluation on 183 Wikipedia biographies shows:

InstructGPT: 42.5% FACTSCORE
ChatGPT: 58.3% FACTSCORE
Perplexity.AI: 71.5% FACTSCORE
Humans: 88.2% FACTSCORE

The automated estimator achieves <2% error rate on a test set of 110 biographies. Retrieval-augmented approaches consistently outperform retrieval-free baselines. Error analysis reveals that models make more errors on rarer entities and information mentioned later in generations, suggesting both pretraining data limitations and error propagation effects.

Scaling analysis of 12 new LLMs shows substantial variance in factual precision: some models like Alpaca (65B) and Vicuna (13B) approach ChatGPT-level performance, while smaller models like Dolly 12B are significantly less factual. Model size correlates with factual precision, though the relationship is not perfectly linear.

Notes

FACTSCORE fills a critical gap in LLM evaluation. While prior work has studied factuality in short-answer QA or summarization, this is the first systematic evaluation of factual precision in long-form biographical generation — a task where error rates compound as the output length grows. The atomic-fact decomposition is elegant and traceable, making the metric interpretable: readers can see exactly which facts the model got wrong. The human evaluation was expensive ($4 per biography at $15–25/hour for Upwork annotators), motivating the automated estimator; however, the estimator's <2% error rate suggests the investment in automation pays off at scale.

One limitation: FACTSCORE measures precision (whether facts are supported) but not recall (whether the generation omits important information). The metric also assumes a single knowledge source (Wikipedia) is sufficient ground truth, which may not hold for more nuanced or domain-specific claims. The scope in this work is also narrow — people biographies — though appendices hint at proof-of-concept extensions to other domains.

The paper has been influential: within months of publication, multiple works cited FACTSCORE for evaluating LLM factuality, confirming that the community needed a robust metric for this task.