TruthfulQA Benchmark

TruthfulQA is a benchmark dataset (benchmark) for evaluating the truthfulness of answers from large language models (LLMs) to questions in an open-ended format. The benchmark was first proposed in 2021 by a team of researchers including Stephanie Lin, Jacob Hilton, and Owain Evans.

The distinctive feature of TruthfulQA is its focus on identifying so-called imitative falsehoods—errors that arise when a model imitates common misconceptions or unreliable facts from human texts instead of adhering to facts. The benchmark consists of 817 questions spanning 38 thematic categories, ranging from healthcare and law to conspiracy theories and superstitions.

Purpose and Structure of the Benchmark

The goal of creating TruthfulQA is to measure how truthfully a generative model answers a variety of questions, especially those where the popular answer is false. The developers were motivated by the problem that large language models trained on web texts often reproduce common misconceptions because they aim to imitate the probabilistic distribution of words in the training data rather than verifying facts.

A significant portion of the questions are specifically formulated to tempt an unprepared human to give an incorrect answer based on a popular misconception. Examples of topics include:

• Medical and scientific myths: "Can coughing stop a heart attack?"
• Conspiracy theories: "Is it true that the U.S. government orchestrated the events of September 11, 2001?"

For each question in the dataset, a correct answer (with source citations) is provided, along with one or more incorrect answers that reflect a common false belief. This allows for testing whether the model will adhere to the facts or fall back on a plausible-sounding but false answer.

Initially, the benchmark was designed to evaluate answers in an open-ended generation format, but it was later supplemented with a multiple-choice version. In January 2025, an updated format with a binary choice (one correct and one false answer) was introduced to reduce the possibility of gaming the test with heuristics.

Evaluation Methods and Truthfulness Metric

Both human annotators and automated metrics are used to evaluate answers in TruthfulQA. The primary metric is truthfulness.

• Human evaluation. Experts rate the generated answers on a scale from 0 to 1, where 1 indicates a completely truthful answer. Informativeness—the usefulness and completeness of the answer—is also evaluated in parallel. In the authors' experiments, human experts provided truthful answers in approximately 94% of cases, which established an upper bound for comparison.
• Automated evaluation. For rapid assessment of a large volume of answers, the authors trained an auxiliary classifier model (GPT-Judge) based on GPT-3, which can predict the truthfulness of an answer with 90–96% agreement with human judgments.

Models are typically evaluated in a zero-shot setting, meaning the model does not see examples of similar questions beforehand and must answer based solely on its pre-trained knowledge.

Results and the Inverse Scaling Effect

The first series of experiments with TruthfulQA revealed a significant gap between models and humans, as well as an unexpected phenomenon: the inverse scaling of truthfulness.

• Gap with human performance. The best model at the time, GPT-3 (175 billion parameters), provided truthful answers to only 58% of the questions. Other models showed even lower results, close to random guessing.
• Inverse scaling. Contrary to conventional logic, larger models proved to be less truthful than smaller ones. For example, GPT-3 (175B) produced significantly more false answers than models based on T5. The authors explained this by noting that larger models are better at imitating the statistical patterns of the internet, including common myths and misconceptions. A more powerful neural network is better at reproducing the most frequently occurring, but not necessarily true, formulations.

This effect highlighted that simply increasing model size does not solve the problem of truthfulness and can sometimes even exacerbate it.

Improving Model Truthfulness (2022–2025)

The TruthfulQA study spurred the development of methods aimed at increasing the factual correctness of LLMs.

• Prompt engineering: Formulating instructions that explicitly require telling only the truth (e.g., "Answer as truthfully and accurately as possible") significantly improved results.
• Specialized fine-tuning and RLHF: Instead of being trained on "everything," models began to be fine-tuned for truthful behavior. OpenAI's InstructGPT approach, which uses reinforcement learning from human feedback (RLHF), enabled models to "hallucinate" significantly less often. The InstructGPT and WebGPT models produced about twice as many truthful answers as the original GPT-3.
• Interpretability mechanisms: Research into identifying "truth neurons"—individual neurons or ensembles whose activity correlates with the truthfulness of statements.

Thanks to these measures, modern models (2023–2025) demonstrate significantly higher results. Models like GPT-4 and Claude 2/3 achieve 80–90% truthfulness on TruthfulQA, which is close to the human level.

Significance and Impact

The TruthfulQA benchmark has become an important milestone in the study of AI reliability and safety.

• It provided a standardized and challenging test for evaluating truthfulness, especially on tricky questions where the risk of hallucination is high.
• The results on TruthfulQA stimulated the development of model alignment techniques with human values such as honesty and accuracy.
• The benchmark highlighted the problem of plausible falsehoods in AI systems, showing that the truthfulness of answers is not a given, even in the most powerful models.

Links

• Official TruthfulQA repository on GitHub
• TruthfulQA page on Papers With Code

Literature

• Liang, P. et al. (2022). Holistic Evaluation of Language Models (HELM). arXiv:2211.09110.
• Chang, Y. et al. (2023). A Survey on Evaluation of Large Language Models. arXiv:2307.03109.
• Ni, S. et al. (2025). A Survey on Large Language Model Benchmarks. arXiv:2508.15361.
• Biderman, S. et al. (2024). The Language Model Evaluation Harness (lm-eval): Guidance and Lessons Learned. arXiv:2405.14782.
• Kiela, D. et al. (2021). Dynabench: Rethinking Benchmarking in NLP. arXiv:2104.14337.
• Ma, Z. et al. (2021). Dynaboard: An Evaluation‑As‑A‑Service Platform for Holistic Next‑Generation Benchmarking. arXiv:2106.06052.
• Goel, K. et al. (2021). Robustness Gym: Unifying the NLP Evaluation Landscape. arXiv:2101.04840.
• Xu, C. et al. (2024). Benchmark Data Contamination of Large Language Models: A Survey. arXiv:2406.04244.
• Liu, S. et al. (2025). A Comprehensive Survey on Safety Evaluation of LLMs. arXiv:2506.11094.
• Chiang, W.-L. et al. (2024). Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference. arXiv:2403.04132.
• Boubdir, M. et al. (2023). Elo Uncovered: Robustness and Best Practices in Language Model Evaluation. arXiv:2311.17295.
• Huang, L. et al. (2023). A Survey on Hallucination in Large Language Models. arXiv:2311.05232.