Elamur Research⌕
☰
home / papers / 2026 / PROBE: PROcess-Based BEnchmark for Hallucination Detection

PROBE: PROcess-Based BEnchmark for Hallucination Detection

Paper · 2026 · Yu Zhang, Peter Belcak, Shizhe Diao et al.

Summary

LLM-as-a-judge setups usually ask a model for a single verdict on whether an output stays faithful to its source. This offers little insight into where detection goes wrong. PROBE treats hallucination detection as a pipeline of four separately scored stages: splitting text into atomic claims, locating evidence in the source, judging whether that evidence is relevant, and pinpointing the hallucinated content. Here, hallucination means unfaithfulness, i.e., content unsupported by the provided context, as in RAG.

The authors built the benchmark from 3,000 Clean Wikipedia articles. GPT-OSS-120B generated summaries, QA pairs and style-transferred rewrites, then injected plausible but unsupported statements at three complexity levels. Frontier models located and voted on evidence to produce claim–evidence pairs.

Every tested model detected hallucinations better when working step by step than with direct prompting. Evidence finding was the main bottleneck, with evidence evaluation also weak. Fine-tuning Llama on the released training data substantially improved the hardest step. PROBE therefore lets researchers diagnose detectors rather than only score them.

Key points

  • PROBE has 12,000 test cases, with 4,000 each for summarization, question answering and style transfer. Each task block contains 1,000 clean samples and 3,000 samples with one, two or three injected hallucinations.
  • Table 1 lists 118,628 annotated claims (25,613 hallucinated, 92,925 truthful), while Table 3 reports 120,404. Hallucinated text makes up about 29.96% of responses overall and is highest in QA (47.42%).
  • Style transfer targets four styles (blog post, lecture notes, FAQ, textbook), each covering roughly a quarter of the samples.
  • The authors position PROBE as the first process-based, claim-level benchmark, in contrast to RAGTruth, HaluBench, HaluEval, FACTS Grounding and FactScore.
  • Stated limitations:
    • English only, non-expert domains.
    • Larger models such as LLaMA-3.1-70B were not fine-tuned.
    • The multi-call pipeline adds latency compared with direct prompting.

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

🎓Chinese University of Hong KongHong Kong · China159 works · 16,154 citations🏢NvidiaSanta Clara · United States37 works · 12,656 citations🏢Nvidia (United Kingdom)Reading · United Kingdom27 works · 11,141 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 works

🧭 Topics 1Research subjects and languages the work is about.Evaluation & benchmarks · 1,341 works

Cite

Cite as Yu Zhang, Peter Belcak, Shizhe Diao, Yonggan Fu, Shaona Ghosh, Morteza Mardani, Eileen Margaret Peters Long, Bei Yu and Pavlo Molchanov. 2026. PROBE: PROcess-Based BEnchmark for Hallucination Detection. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.2099

@inproceedings{zhang2026probe,
  title = {{PROBE: PROcess-Based BEnchmark for Hallucination Detection}},
  author = {Yu Zhang, Peter Belcak, Shizhe Diao, Yonggan Fu, Shaona Ghosh, Morteza Mardani, Eileen Margaret Peters Long, Bei Yu and Pavlo Molchanov},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.2099},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.2099}
}

References

27 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. AI@Meta. 2024. Llama 3 model card. find
  2. Shayan Ali Akbar, Md Mosharaf Hossain, Tess Wood, Si-Chi Chin, Erica M Salinas, Victor Alvarez, and Erwin Cornejo. 2024. Hallumeasure: Fine-grained hallucination measurement using chain-of-thought reasoning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, p… find
  3. Dean Allemang and Juan Sequeda. 2024. Increasing the llm accuracy for question answering: Ontologies to the rescue! arXiv preprint arXiv:2405.11706. — [arXiv:2405.11706](https://arxiv.org/abs/2405.11706) find
  4. Anthropic. 2025. Introducing claude sonnet 4.5. find
  5. Sabine Foly, Jingshu Liu, Jean-Gabriel Barthelemy, Gaëtan Caillaut, Raheel Qader, Mariam Nakhle, and Arezki Sadoune. 2025. articles. find
  6. Clean-wikipedia-english-Xiangkun Hu, Dongyu Ru, Lin Qiu, Qipeng Guo, Tianhang Zhang, Yang Xu, Yun Luo, Pengfei Liu, Yue Zhang, and Zheng Zhang. 2024. Refchecker: Reference-based fine-grained hallucination checker and benchmark for large language models. arXiv preprint arXiv:2405.14486. — [arXiv:2405.14486](https://arxiv.org/abs/2405.14486) find
  7. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276.
  8. Alon Jacovi, Andrew Wang, Chris Alberti, Connie Tao, Jon Lipovetz, Kate Olszewska, Lukas Haas, Michelle Liu, Nate Keating, Adam Bloniarz, and 1 others. 2025. The facts grounding leaderboard: Benchmarking llms’ ability to ground responses to long-form input. arXiv preprint arXiv:2501.03200.
  9. Diederik P Kingma. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.
  10. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, and 1 others. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33:9459–…
  11. Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2023. Halueval: A large-scale hallucination evaluation benchmark for large language models. arXiv preprint arXiv:2305.11747.
  12. Junyi Li, Tianyi Tang, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2024. Pre-trained language models for text generation: A survey. ACM Computing Surveys, 56(9):1–39. find
  13. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Let’s verify step by step. In The Twelfth International Conference on Learning Representations.
  14. Zicheng Lin, Zhibin Gou, Tian Liang, Ruilin Luo, Haowei Liu, and Yujiu Yang. 2024. Criticbench: Benchmarking llms for critique-correct reasoning. arXiv preprint arXiv:2402.14809. find
  15. Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 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 L…
  16. Abhika Mishra, Akari Asai, Vidhisha Balachandran, Yizhong Wang, Graham Neubig, Yulia Tsvetkov, and Hannaneh Hajishirzi. 2024. Fine-grained hallucination detection and editing for language models. arXiv preprint arXiv:2401.06855.
  17. Zihao Zhou, Shudong Liu, Maizhen Ning, Wei Liu, Jindong Wang, Derek F Wong, Xiaowei Huang, Qiufeng Wang, and Kaizhu Huang. 2024. Is your model really a good math reasoner? evaluating mathematical reasoning with checklist. arXiv preprint arXiv:2407.08733. — [arXiv:2407.08733](https://arxiv.org/abs/2407.08733) find
  18. Mistral.ai. 2024. Mixtral-8x22b. find
  19. Cheng Niu, Yuanhao Wu, Juno Zhu, Siliang Xu, Kashun Shum, Randy Zhong, Juntong Song, and Tong Zhang. 2023. Ragtruth: A hallucination corpus for developing trustworthy retrieval-augmented language models. arXiv preprint arXiv:2401.00396.
  20. OpenAI. 2025. gpt-oss-120b & gpt-oss-20b model card. Preprint, arXiv:2508.10925.
  21. Selvan Sunitha Ravi, Bartosz Mielczarek, Anand Kannappan, Douwe Kiela, and Rebecca Qian. 2024. Lynx: An open source hallucination evaluation model. arXiv preprint arXiv:2407.08488. — [arXiv:2407.08488](https://arxiv.org/abs/2407.08488) find
  22. Juntong Song, Xingguang Wang, Juno Zhu, Yuanhao Wu, Xuxin Cheng, Randy Zhong, and Cheng Niu. 2024. Rag-hat: A hallucination-aware tuning pipeline for llm in retrieval-augmented generation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, page… find
  23. Yingjia Wan, Haochen Tan, Xiao Zhu, Xinyu Zhou, Zhiwei Li, Qingsong Lv, Changxuan Sun, Jiaqi Zeng, Yi Xu, Jianqiao Lu, and 1 others. 2025. Fastfact: Faster, stronger long-form factuality evaluations in llms. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 23814–23854. find
  24. Jerry Wei, Chengrun Yang, Xinying Song, Yifeng Lu, Nathan Hu, Jie Huang, Dustin Tran, Daiyi Peng, Ruibo Liu, Da Huang, and 1 others. 2024. Long-form factuality in large language models. Advances in Neural Information Processing Systems, 37:80756–80827.
  25. Yang Zhang, Hanlei Jin, Dan Meng, Jun Wang, and Jinghua Tan. 2024. A comprehensive survey on process-oriented automatic text summarization with exploration of llm-based methods. arXiv preprint arXiv:2403.02901. — [arXiv:2403.02901](https://arxiv.org/abs/2403.02901) find
  26. Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien-Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, and 1 others. 2023. Pytorch fsdp: experiences on scaling fully sharded data parallel. arXiv preprint arXiv:2304.11277. find
  27. Chujie Zheng, Zhenru Zhang, Beichen Zhang, Runji Lin, Keming Lu, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. 2024. Processbench: Identifying process errors in mathematical reasoning. arXiv preprint arXiv:2412.06559.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-probe-process-based-benchmark-hallucination/