Elamur Research⌕
☰
home / papers / 2026 / KGHaluBench: A Knowledge Graph-Based Hallucination Benchmark for Evaluating the Breadth and Depth of LLM Knowledge

KGHaluBench: A Knowledge Graph-Based Hallucination Benchmark for Evaluating the Breadth and Depth of LLM Knowledge

Paper · 2026 · Alex Robertson, Huizhi Liang, Mahbub Gani et al.

Summary

Hallucination benchmarks often rely on fixed question sets and short or multiple-choice formats. These go stale, cover a narrow slice of knowledge, and rarely test how deeply a model knows a topic. KGHaluBench addresses this by building questions dynamically from Wikidata. It samples random entities, balancing very common, common and uncommon entity types. For each entity, it selects three relations from the entity's one-hop subgraph and merges them into one open-ended compound question. Question difficulty is estimated statistically from popularity signals, which counters the bias toward well-known entities.

Answers are checked in stages. First, abstentions are detected. Next, the answer is compared with the entity's Wikipedia description to confirm the model is discussing the right entity. Finally, individual claims are verified against KG triples. The authors evaluate 25 LLMs, both open and proprietary, with three new metrics:

  • difficulty-weighted accuracy (W_a)
  • breadth-of-knowledge hallucination rate (Halu_BOK)
  • depth-of-knowledge hallucination rate (Halu_DOK)

The benchmark proves hard even for frontier models. Larger models mostly identify the entity correctly but still get specific facts wrong. The resource is public, but its dependence on Wikidata means non-English entities are under-represented.

Key points

  • The best model, GPT-5, reaches only 65.60% W_a. Entity popularity is estimated from page views (2017–2025), site links, KG neighbours and the number of external database IDs.
  • The automated verifier agrees with human judgment 79.19% of the time at the entity level and 87.74% at the fact level.
  • Moving from 8–32B models to large proprietary ones, Halu_BOK drops from 54.75% to 11.91%. Halu_DOK falls only by about 24.59%. Even for strong models, roughly 39.74% of answers still contain exactly one hallucinated fact.
  • Abstention lowers hallucination: Llama-3.1-8B scores 30.00% versus 58.67% for Gemma-2-9B despite equal alignment. However, GPT-oss-20B (55.93%), GPT-5-Mini (49.67%) and Claude Sonnet 4 (49.40%) refuse about half or more of the questions.
  • Code is available at github.com/c0037654Newcastle/KGHaluBench. The conclusion cites 27 models, which conflicts with the 25 stated elsewhere in the paper.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Newcastle UniversityNewcastle upon Tyne · United Kingdom18 works · 619 citations

Also on the byline: Sage.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 works

🧭 Topics 3Research subjects and languages the work is about.Language models & pretraining · 3,320 worksEvaluation & benchmarks · 1,220 worksKnowledge graphs & ontologies · 192 works

Cite

Cite as Alex Robertson, Huizhi Liang, Mahbub Gani, Rohit Kumar and Srijith Rajamohan. 2026. KGHaluBench: A Knowledge Graph-Based Hallucination Benchmark for Evaluating the Breadth and Depth of LLM Knowledge. Findings of the Association for Computational Linguistics: EACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-eacl.206

@inproceedings{robertson2026kghalubench,
  title = {{KGHaluBench: A Knowledge Graph-Based Hallucination Benchmark for Evaluating the Breadth and Depth of LLM Knowledge}},
  author = {Alex Robertson, Huizhi Liang, Mahbub Gani, Rohit Kumar and Srijith Rajamohan},
  booktitle = {Findings of the Association for Computational Linguistics: EACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-eacl.206},
  url = {https://doi.org/10.18653/v1/2026.findings-eacl.206}
}

References

46 references, 28 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Sören Auer, Christian Bizer, Georgi Kobilarov, and 1 others. 2007. Dbpedia: a nucleus for a web of open data. In Proceedings of the 6th International The Semantic Web and 2nd Asian Conference on Asian Semantic Web Conference, ISWC’07/ASWC’07, page 722–735, Berlin, Heidelberg. Springer-Verlag. find
  2. Yejin Bang, Ziwei Ji, Alan Schelten, and 1 others. 2025. HalluLens: LLM hallucination benchmark. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 24128–24156, Vienna, Austria. Association for Computational Linguistics.
  3. Jonathan Berant, Andrew K. Chou, Roy Frostig, and 1 others. 2013. Semantic parsing on freebase from question-answer pairs. In Conference on Empirical Methods in Natural Language Processing.
  4. Akhiad Bercovich, Itay Levy, Izik Golan, and 1 others. 2025. Llama-nemotron: Efficient reasoning models. Preprint, arXiv:2505.00949.
  5. Antoine Bordes, Nicolas Usunier, Sumit Chopra, and 1 others. 2015. Large-scale simple question answering with memory networks. Preprint, arXiv:1506.02075. find
  6. Preetam Prabhu Srikar Dammu, Himanshu Naidu, and Chirag Shah. 2025. Dynamic-kgqa: A scalable framework for generating adaptive question answering datasets. Preprint, arXiv:2503.05049. — [arXiv:2503.05049](https://arxiv.org/abs/2503.05049) find
  7. DeepSeek-AI, Daya Guo, Dejian Yang, and 1 others. 2025a. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. Preprint, arXiv:2501.12948.
  8. DeepSeek-AI, Aixin Liu, Bei Feng, and 1 others. 2025b. Deepseek-v3 technical report. Preprint, arXiv:2412.19437. find
  9. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, and 1 others. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.
  10. Dan Hendrycks, Collin Burns, Steven Basart, and 1 others. 2021. Measuring massive multitask language understanding. Preprint, arXiv:2009.03300.
  11. Kelvin Jiang, Dekun Wu, and Hui Jiang. 2019. FreebaseQA: A new factoid QA data set matching trivia-style question-answer pairs with Freebase. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (L… find
  12. Mandar Joshi, Eunsol Choi, Daniel Weld, and 1 others. 2017. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1601–1611, Vancouver, Canada…
  13. Jean Kaddour, Joshua Harris, Maximilian Mozes, Herbie Bradley, Roberta Raileanu, and Robert McHardy. 2023. Challenges and applications of large language models. Preprint, arXiv:2307.10169.
  14. Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, and Edwin Zhang. 2025. Why language models hallucinate. Preprint, arXiv:2509.04664. — [arXiv:2509.04664](https://arxiv.org/abs/2509.04664) find
  15. Jungo Kasai, Keisuke Sakaguchi, Yoichi Takahashi, and 1 others. 2024. Realtime qa: What’s the answer right now? Preprint, arXiv:2207.13332.
  16. Felix Busch, Lena Hoffmann, Christopher Rueger, and 1 others. 2025. Current applications and challenges in large language models for patient care: a systematic review. Communications Medicine, 5. find
  17. Wojciech Kryscinski, Bryan McCann, Caiming Xiong, and 1 others. 2020. Evaluating the factual consistency of abstractive text summarization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9332–9346, Online. Association for Computational Lingui…
  18. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, and Ice Pasupat. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. Preprint, arXiv:2507.06261.
  19. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, and 1 others. 2019. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:452–466.
  20. Vladimir I. Levenshtein. 1965. Binary codes capable of correcting deletions, insertions, and reversals. Soviet physics. Doklady, 10:707–710. find
  21. Junyi Li, Xiaoxue Cheng, Xin Zhao, and 1 others. 2023. HaluEval: A large-scale hallucination evaluation benchmark for large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6449–6464, Singapore. Association for Computational Linguistic…
  22. Yinheng Li, Shaofei Wang, Han Ding, and Hang Chen. 2024. Large language models in finance: A survey. Preprint, arXiv:2311.10723. — [arXiv:2311.10723](https://arxiv.org/abs/2311.10723) find
  23. Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. TruthfulQA: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214–3252, Dublin, Ireland. Association for Computational Linguistics…
  24. Xiaoze Liu, Feijie Wu, Tianyang Xu, and 1 others. 2024. Evaluating the factuality of large language models using large-scale knowledge graphs. Preprint, arXiv:2404.00942. — [arXiv:2404.00942](https://arxiv.org/abs/2404.00942) find
  25. Sewon Min, Kalpesh Krishna, Xinxi Lyu, and 1 others. 2023. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12076–12100, Singapore. Association for Computation…
  26. Shervin Minaee, Tomas Mikolov, Narjes Nikzad, Meysam Chenaghlu, Richard Socher, Xavier Amatriain, and Jianfeng Gao. 2025. Large language models: A survey. Preprint, arXiv:2402.06196.
  27. Niklas Muennighoff, Nouamane Tazi, Loic Magne, and 1 others. 2023. MTEB: Massive text embedding benchmark. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2014–2037, Dubrovnik, Croatia. Association for Computational Linguistics.
  28. Pranav Narayanan Venkit, Tatiana Chakravorti, Vipul Gupta, and 1 others. 2024. An audit on the perspectives and challenges of hallucinations in NLP. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 6528–6548, Miami, Florida, USA. Association for Comput… find
  29. OpenAI. 2024. Gpt-4 technical report. arXiv:2303.08774.
  30. Preprint, Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram, and 1 others. 2024. jina-embeddings-v3: Multilingual embeddings with task lora. Preprint, arXiv:2409.10173. find
  31. Kai Sun, Yifan Ethan Xu, Hanwen Zha, and 1 others. 2024. Head-to-tail: How knowledgeable are large language models (llms)? a.k.a. will llms replace knowledge graphs? Preprint, arXiv:2308.10168. find
  32. Alon Talmor and Jonathan Berant. 2018. The web as a knowledge-base for answering complex questions. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 641–651, New Orleans, L…
  33. Gemma Team, Morgane Riviere, Shreya Pathak, and 1 others. 2024. Gemma 2: Improving open language models at a practical size. Preprint, arXiv:2408.00118. find
  34. Kimi Team. 2025b. Kimi k2: Open agentic intelligence. Preprint, arXiv:2507.20534.
  35. James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and 1 others. 2018. FEVER: a large-scale dataset for fact extraction and VERification. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume…
  36. Denny Vrandečić and Markus Krötzsch. 2014. Wikidata: a free collaborative knowledgebase. Commun. ACM, 57(10):78–85.
  37. Tu Vu, Mohit Iyyer, Xuezhi Wang, and 1 others. 2023. Freshllms: Refreshing large language models with search engine augmentation. Preprint, arXiv:2310.03214.
  38. Jason Wei, Nguyen Karina, Hyung Won Chung, and 1 others. 2024. Measuring short-form factuality in large language models. Preprint, arXiv:2411.04368.
  39. Wikidata. 2025. Wikidata:statistics - wikidata. find
  40. Qwen, An Yang, Baosong Yang, and 1 others. 2025. Qwen2.5 technical report. Preprint, arXiv:2412.15115. find
  41. A B M Ashikur Rahman, Saeed Anwar, Muhammad Usman, and Ajmal Mian. 2024. Defan: Definitive answer dataset for llms hallucination evaluation. Preprint, arXiv:2406.09155. — [arXiv:2406.09155](https://arxiv.org/abs/2406.09155) find
  42. Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. 2024. Hallucination is inevitable: An innate limitation of large language models. ArXiv, abs/2401.11817.
  43. An Yang, Anfeng Li, Baosong Yang, and 1 others. 2025. Qwen3 technical report. Preprint, arXiv:2505.09388.
  44. Zhilin Yang, Peng Qi, Saizheng Zhang, and 1 others. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369–2380, Brussels, Belgium. Association for Computational Linguistic…
  45. Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, and 1 others. 2023. Siren’s song in the ai ocean: A survey on hallucination in large language models. Preprint, arXiv:2309.01219.
  46. Yanxu Zhu, Jinlin Xiao, Yuhang Wang, and 1 others. 2024. Kg-fpq: Evaluating factuality hallucination in llms with knowledge graph-based false premise questions. Preprint, arXiv:2407.05868. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/robertson-2026-kghalubench-knowledge-graph-based-hallucination/