Survey of Hallucination in Natural Language Generation
Open accessThe DOI matches its Crossref record; 355 of the 360 references it cites were found in a registry or an index, and 4 are not checked yet; 224 of the 256 sources its claims cite were checked for support.
Abstract
Natural Language Generation (NLG) has improved exponentially in recent years thanks to the development of sequence-to-sequence deep learning technologies such as Transformer-based language models. This advancement has led to more fluent and coherent NLG, leading to improved development in downstream tasks such as abstractive summarization, dialogue generation, and data-to-text generation. However, it is also apparent that deep learning based generation is prone to hallucinate unintended text, which degrades the system performance and fails to meet user expectations in many real-world scenarios. To address this issue, many studies have been presented in measuring and mitigating hallucinated texts, but these have never been reviewed in a comprehensive manner before. In this survey, we thus provide a broad overview of the research progress and challenges in the hallucination problem in NLG. The survey is organized into two parts: (1) a general overview of metrics, mitigation methods, and future directions, and (2) an overview of task-specific research progress on hallucinations in the following downstream tasks, namely abstractive summarization, dialogue generation, generative question answering, data-to-text generation, and machine translation. This survey serves to facilitate collaborative efforts among researchers in tackling the challenge of hallucinated texts in NLG.
Verified 2026-11-19 · 4 references not checked yet
Overview
CoChat Summary
TL;DR
This survey provides a broad overview of hallucination in natural language generation, covering definitions, contributors, metrics, mitigation methods, and task-specific progress across summarization, dialogue, GQA, data-to-text, machine translation, vision-language generation, and LLMs. It distinguishes intrinsic from extrinsic hallucination and argues that hallucination must be treated differently across tasks and sources.
Key findings
- The survey categorizes hallucination into intrinsic hallucination, where generated output contradicts the source, and extrinsic hallucination, where output cannot be verified from the source; extrinsic hallucination is not always erroneous because it may come from factually correct external information.
- It reports prevalence examples from cited work: 25% of summaries generated by state-of-the-art abstractive summarization systems have hallucinated content; 62% of first sentences in WIKIBIO have additional information not stated in the corresponding infobox; and 74% of samples in the QMSum dataset consist of inconsistent facts.
- It reviews hallucination metrics as statistical, model-based (IE-based, QA-based, NLI-based, faithfulness classification, LM-based), and human evaluation, and mitigation methods as data-related and modeling/inference methods including architecture changes, planning/sketching, RL, multi-task learning, controllable generation, and post-processing.
- For LLMs, it notes that RLHF reduced the hallucination rate from 41% to 21% in closed-domain tasks for InstructGPT, and discusses reference-dependent and reference-free hallucination metrics plus mitigation via data refinement, RAG, safety fine-tuning, model editing, decoding methods, chain-of-thought variants, post-processing, and ensembling.
Methods
- Literature survey design: the paper reviews existing research on hallucination in NLG rather than conducting new experiments or primary data collection.
- Coverage: general NLG hallucination topics (definitions, contributors, metrics, mitigation, future directions) plus task-specific sections on abstractive summarization, dialogue generation, generative question answering, data-to-text generation, neural machine translation, vision-language generation, and large language models.
- Analysis: synthesizes and compares definitions, metrics, mitigation approaches, and future directions across tasks, with examples in tables and figures; the LLM section was updated in Jan 2024.
Limitations
- The paper states that GQA has no standard definition or categorization of hallucination and no automatic metric specifically for hallucination in GQA; it also notes that open-domain dialogue lacks a standard hallucination metric and that VL hallucination research is still in its early stage.
- The paper states that extrinsic hallucination has been somewhat overlooked and is more challenging to reduce, and that most existing mitigation methods do not distinguish between intrinsic and extrinsic hallucination.
- Evident from the text: as a survey, it does not provide new empirical validation of metrics or mitigation methods, and its LLM coverage is explicitly updated only through Jan 2024, so later work is not included.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures
Verification
Claims the paper makes
124 Valid · 26 Not valid · 19 Not checked · 135 Own results and interpretations
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“Furthermore, perturbation, where the insertions of most common tokens are placed at the beginning of the sentence, seems to be the most successful in hallucination mitigation.”
What the sources say
- Not checked — no full text of the sourcecites [133] [133] Katherine Lee, Orhan Firat, Ashish Agarwal, Clara Fannjiang, and David Sussillo. 2019. Hallucinations in Neural Machine Translation. ICLR (2019). -
- Disputedcites [215] [215] Vikas Raunak, Arul Menezes, and Marcin Junczys-Dowmunt. 2021. The Curious Case of Hallucinations in Neural Machine Translation. (4 2021), 1172-1183. https://arxiv.org/abs/2104.06683v1 - sourceExperimental — accuracy not yet measured
In the passage describing the perturbation procedure (page unavailable), the paper says that inserting common tokens at the beginning is used to generate hallucinations, not mitigate them. Its prevention section instead discusses approaches such as robust learning and corpus-level filtering. (LLM, conf=0.92)
Checked by a reasoning model — may be wrong.
hedgedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“While there are works on the visual extension of faithfulness-oriented RLHF [303], these essentially involve labeling more data, which is expensive and difficult to scale.”
What the sources say
- Disputedcites [303] [303] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, and Tat-Seng Chua. 2023. RLHF-V: Towards Trustworthy MLLMs via Behavior Alignment from Fine-grained Correctional Human Feedback. CoRR abs/2312.00849 (2023). https://doi.org/10.48550/ARXIV.2312.00849 arXiv:2312.00849 - sourceExperimental — accuracy not yet measured
RLHF-V does use human annotation for visual-model alignment, but it makes the contrary point about needing more labeled data. The Fine-grained Correctional Human Preference Collection section says its corrections improve data efficiency, and Section 4.3 reports comparable hallucination rates using only 200 preference examples versus an order of magnitude more ranking-labeled examples. Our copy does not establish a separate claim about annotation cost or scalability. (LLM, conf=0.88)
Checked by a reasoning model — may be wrong.
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Disputed
Every source the paper cites for this claim was checked and supports it.
In the paper
“Nevertheless, as illustrated by Falke et al. [66], off-the-shelf NLI models tend to transfer poorly to the abstractive summarization task.”
What the sources say
- Supportedcites [66] [66] Tobias Falke, Leonardo FR Ribeiro, Prasetya Ajie Utama, Ido Dagan, and Iryna Gurevych. 2019. Ranking Generated Summaries by Correctness: An Interesting but Challenging Application for Natural Language Inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics . 2214-2220. - source
The Introduction states that models trained on NLI datasets transfer poorly to the paper’s downstream task of ranking generated summaries by correctness. The Conclusions likewise say that out-of-the-box NLI models do not perform well on that task. (LLM, conf=0.94)
- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Such wrong association results in generation with facts mixed up between two similar entities [60, 233].”
What the sources say
- Partly supportedcites [60] [60] Nouha Dziri, Andrea Madotto, Osmar Zaiane, and Avishek Joey Bose. 2021. Neural Path Hunter: Reducing Hallucination in Dialogue Systems via Path Grounding. EMNLP (2021). - sourceExperimental — accuracy not yet measured
The GPT2-KG passage describes an incorrect association between entities, and G Error Analysis gives an example of a generated response assigning a book to the wrong author. These support the claim that facts can be mixed up, but the paper does not say the entities involved are similar. The supplied passages do not provide page numbers. (LLM, conf=0.83)
- Supportedcites [233] [233] Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. 2021. Retrieval Augmentation Reduces Hallucination in Conversation. EMNLP (2021). - source
The Introduction states that language models often mix up facts between two similar entities when generating factually incorrect statements. This directly supports the cited description of the resulting error. (LLM, conf=0.92)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“As a result, the target reference may contain information that cannot be supported by the source [198, 268].”
What the sources say
- Supportedcites [198] [198] Ankur Parikh, Xuezhi Wang, Sebastian Gehrmann, Manaal Faruqui, Bhuwan Dhingra, Diyi Yang, and Dipanjan Das. 2020. ToTTo: A Controlled Table-To-Text Generation Dataset. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) . 1173-1186. - source
The Dataset section says that reference targets selected from web sentences may contain information that cannot be inferred from the source. The Data Annotation Process section also describes deleting phrases unsupported by table cells. (LLM, conf=0.93)
- Supportedcites [268] [268] Hongmin Wang. 2019. Revisiting Challenges in Data-to-Text Generation with Fact Grounding. In Proceedings of the 12th International Conference on Natural Language Generation . 311-322. - source
The Introduction reports that about 40% of human-written game-summary content cannot be directly mapped to the input boxscore records and draws on information beyond those inputs. Section 2.1 likewise describes reference content not licensed by the boxscore table. This supports the claim that a target reference may contain information the source does not support. (LLM, conf=0.93)
hedgedstates a possible cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
117 more claims · 74 with their cited sources checked
Create a free account to see all 304 claims and to verify the 2 references still open.
- Borderline: one of the tests for a substantial claim is unclear. 1
- It may state two separate claims as one. 90
- Probably not a substantial claim. 142
- The quote only partly states this claim. 1
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Hallucination can lead to serious consequences such as incorrect or misleading code, security risks, and ethical issues [186].”
From A survey on large language model based autonomous agents (2024) · cites this paper as [186]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“citation fabrication constitutes a form of extrinsic hallucination in the absence of retrieval augmentation, since the model generates references not grounded in any provided source material [4].”
From How LLMs Cite and Why It Matters: A Cross-Model Audit of Reference Fabrication in AI-Assisted Academic Writing and Methods to Detect Phantom Citations (2026) · cites this paper as [4]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Recent taxonomic work distinguishes at least seven cases [15,18].”
From Hallucination as epistemic pollution: epistemic injustice and credibility in futures-oriented science education (2026) · cites this paper as [18]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“LLMs struggle to generate accurate and up-to-date responses on current topics, due to outdated knowledge, lack of domain-specific expertise, or hallucination13–17.”
From Leveraging long context in retrieval augmented language models for medical question answering (2025) · cites this paper as [16]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Fine-tuning can enrich LLMs with domain knowledge, enhance their specific ability, improve the fairness and reliability of the outputs, and prevent certain damage caused by hallucination [26].”
From FederatedScope-LLM: A Comprehensive Package for Fine-tuning Large Language Models in Federated Learning (2024) · cites this paper as [26]
Experimental — accuracy not yet measured
41 more claims about this paper
Create a free account to see every claim other papers make about this one.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Extrinsic Object Hallucination
“CHAIR (Caption Hallucination Assessment with Image Relevance) metric, which calculates what proportion of object words generated are actually in the image according to the ground truth captions.”
§ Extrinsic Object Hallucination
“CHAIR 𝑖 measures per-instance object hallucination, i.e. what fraction of object instances in each generated caption are hallucinated.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 7.1 Hallucination Definition in Abstractive Summarization
“For instance, in Table 1, given the input article shown in the caption, an example of intrinsic hallucination is ' The Ebola vaccine was rejected by the FDA in 2019, '”
§ 2.1 Categorization
“in the abstractive summarization task from Table 1, the generated summary ' The first Ebola vaccine was approved in 2021 ' contradicts the source content”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Survey of Hallucination in Natural Language Generation
“The survey is organized into two parts: (1) a general overview of metrics, mitigation methods, and future directions; (2) an overview of task-specific research progress on hallucinations in the following downstream tasks, namely abstractive summarization, dialogue generation, generative question answering, data-to-text generation, machine translation, and visual-language generation; and (3) hallucinations in large language models (LLMs)”
§ 1 INTRODUCTION
“The third part discusses this phenomenon in LLMs in Section 13.”
Citations
Cites
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325 · cited 3×
“Recently, the rapid development of NLG has captured the imagination of many thanks to the advances in deep learning technologies, especially Transformer [189]-based models like BERT [29], BART [100], GPT-2 [149], and GPT-3 [16].”
BackgroundAn Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. (2020) · arXiv · doi:10.48550/arxiv.2010.11929 · Cited by 47,472
Language Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372
Sequence to Sequence Learning with Neural Networks
Sutskever, I., Vinyals, O., & Le, Q. V. (2014) · arXiv · doi:10.48550/arxiv.1409.3215 · Cited by 17,766
Advances in Neural Information Processing Systems 29
Teymur, O., Zygalakis, K., & Calderhead, B. (2016) · Cited by 13,887
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2019) · arXiv · doi:10.48550/arxiv.1910.10683 · Cited by 10,929 · cited 1×
“This work investigates the performances of GPT-3 [13], GPT-Neo/J [141], GPT-2 [105] and a T5-based model [106].”
BackgroundAdvances in Neural Information Processing Systems 14: Proceedings of the 2001 Conference
(2002) · The MIT Press eBooks · doi:10.7551/mitpress/1120.001.0001 · Cited by 8,890
Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., & Zettlemoyer, L. (2020) · Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/2020.acl-main.703 · Cited by 8,821 · cited 1×
“Recently, the development of NLG has conquered the imagination of many thanks to the advances in deep learning technologies, especially the Transformer [159]-based models like BERT [24], BART [83], GPT-2 [127], and GPT3 [13].”
BackgroundLewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., & Zettlemoyer, L. (2019) · arXiv · doi:10.48550/arxiv.1910.13461 · Cited by 8,816 · cited 1×
“Recently, the development of NLG has conquered the imagination of many thanks to the advances in deep learning technologies, especially the Transformer [159]-based models like BERT [24], BART [83], GPT-2 [127], and GPT3 [13].”
Background
Research works printed but not linked: 16
- [4] Joshua Albrecht and Rebecca Hwa. 2007. A Re-examination of Machine Learning Approaches for Sentence-Level MT Evaluation. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics . Association for Computational Linguistics, Prague, Czech Republic, 880-887. https://aclanthology.org/P07-1111 -
- [16] Andrzej Białecki, Robert Muir, Grant Ingersoll, and Lucid Imagination. 2012. Apache lucene 4. In SIGIR 2012 workshop on open source information retrieval . 17. -
- [53] Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. 2023. Improving Factuality and Reasoning in Language Models through Multiagent Debate. In Forty-first International Conference on Machine Learning . -
- [121] Philipp Koehn and Rebecca Knowles. 2017. Six Challenges for Neural Machine Translation. In First Workshop on Neural Machine Translation . Association for Computational Linguistics, 28-39. -
- [145] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation. In ICML . -
- [149] Tian Li, Ahmad Beirami, Maziar Sanjabi, and Virginia Smith. 2020. Tilted Empirical Risk Minimization. In International Conference on Learning Representations . -
- [164] Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. 2018. Neural Baby Talk. In CVPR . -
- [165] Yukun Ma, Khanh Linh Nguyen, Frank Z Xing, and Erik Cambria. 2020. A Survey on Empathetic Dialogue Systems. Information Fusion 64 (2020), 50-70. -
- [175] Marianna J. Martindale, Marine Carpuat, Kevin Duh, and Paul McNamee. 2019. Identifying Fluently Inadequate Output in Neural and Statistical Machine Translation. In MTSummit . -
- [179] Kathleen McKeown. 1992. Text Generation . Cambridge University Press. -
- [209] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language Models are Unsupervised Multitask Learners. OpenAI Blog 1, 8 (2019), 9. -
- [238] Hui Su, Xiaoyu Shen, Sanqiang Zhao, Zhou Xiao, Pengwei Hu, Cheng Niu, and Jie Zhou. 2020. Diversifying Dialogue Generation with Non-Conversational Text. In 58th Annual Meeting of the Association for Computational Linguistics . ACL, 7087-7097. -
- [256] Zhaopeng Tu, Yang Liu, Lifeng Shang, Xiaohua Liu, and Hang Li. 2017. Neural Machine Translation with Reconstruction. In Thirty-First AAAI Conference on Artificial Intelligence . -
- [261] Neeraj Varshney, Wenlin Yao, Hongming Zhang, Jianshu Chen, and Dong Yu. 2023. A Stitch in Time Saves Nine: Detecting and Mitigating Hallucinations of LLMs by Validating Low-Confidence Generation. CoRR abs/2307.03987 (2023). https://doi.org/10.48550/arXiv.2307.03987 arXiv:2307.03987 - · arXiv · doi:10.48550/arxiv.2307.03987 · Paper page
- [285] Zeqiu Wu, Michel Galley, Chris Brockett, Yizhe Zhang, Xiang Gao, Chris Quirk, Rik Koncel-Kedziorski, Jianfeng Gao, Hannaneh Hajishirzi, Mari Ostendorf, et al. 2021. A Controllable Model of Grounded Response Generation. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 35. 14085-14093. -
- [291] Xinnuo Xu, Ondřej Dušek, Verena Rieser, and Ioannis Konstas. 2021. AGGGEN: Ordering and Aggregating while Generating. roceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL2021) (2021). -
Cited by
Items 4,301–4,400 of 4,844
Alqahtani, M. Q., & Abumelha, M. (2026) · Journal of King Saud University - Computer and Information Sciences · doi:10.1007/s44443-026-01068-9
“Hallucination remains a known challenge across all large language models (LLMs), where the model may generate inaccurate or fabricated content (Ji et al. 2023; Wu et al. 2025).”
Agent-assisted planning and optimization for load-following operation of pressurized water reactorsPreprint
Shi, H., Zhao, C., Niu, Z., Zhang, S., Wang, X., & He, M. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7308319
Brendel, B., Lichtenberg, S., Hildebrandt, F., & Dennis, A. R. (2026) · Journal of Management Information Systems · doi:10.1080/07421222.2026.2692275
“Because humans are inconsistent [69, 96, 112] and frequently err [26], errors are more normal.”
InfluentialDecker, R. (2026) · Springer eBooks · doi:10.1007/978-3-658-51525-6_69
How Reflection Enhances Task Factuality in the Use of Large Language Models
Tarafdar, M., Adam, M., & Nguyen, L. T. (2026) · Journal of Management Information Systems · doi:10.1080/07421222.2026.2692283 · Cited by 2
“…more accurate and relevant information retrieval, error-prone and halluci-nated outputs are still possible because of the inherently probabilistic nature of LLMs’ computational processing that can result in fabrication or misattribution [62] and because LLMs lack deep contextual information [20].”
Akintande, O. J. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7295938
An LLM and Retrieval Pipeline for Reproductive-Health Misinformation and Stigma
Behnamian, S., Shahbazi, Z., Fogh, F., Baghestani, B., Khani, A., & Rostami, A. D. (2026) · Reproductive Medicine · doi:10.3390/reprodmed7030042
Research on Simplified Design of Interface Interaction Process Based on TRIZ and Generative AI
Li, M., Liu, S., Wu, B., Zhao, X., & Ning, Z. (2026) · International Journal of Human-Computer Interaction · doi:10.1080/10447318.2026.2705084
di Cicco, P., Bennar, W., Volet, C., Puricel, S., Togni, M., Cook, S., & Garín, D. (2026) · Cardiovascular Medicine · doi:10.3390/cardiovascmed29030031
“This predictive mechanism explains the hallucination of precise-seeming percentages and the models’ tendency to express unwarranted confidence in objectively incorrect calculations [17].”
Crafting Effective Prompts: Thinking with Machines Not Just Talking to Them
Sinha, E., Chandra, S., & Waghmare, C. D. (2026) · Emerald Group Publishing Limited eBooks · doi:10.1108/978-1-83742-179-420261002
Xie, Q., Fang, Y., & Hu, Y. (2026) · Frontiers in Immunology · doi:10.3389/fimmu.2026.1937268
“LLMs generate fl uent output that can be factually wrong, and different models — trained on different data and objectives — may read the same sentence differently (38).”
Tan, C. H., & Ding, Q. (2026) · FinTech · doi:10.3390/fintech5030071
Na, K., & Jeong, Y. (2026) · The Journal of Academic Librarianship · doi:10.1016/j.acalib.2026.103328
Yoo, Y., & Kim, S. (2026) · INFORMS Transactions on Education · doi:10.1287/ited.2025.0175
Li, Y., Jin, G., Jiang, X., et al. (2026) · IEEE Access · doi:10.1109/access.2026.3724626
“However, deploying LLMs directly for agronomic recommendations is risky due to hallucinations, plausible but factually incorrect outputs, especially in domains requiring regulatory compliance [22].”
Translate, Search, or Answer: Cost-Aware Cross-Lingual Retrieval for Kazakh Small Language Models
Maxutov, A., Medeu, N., Albrekht, V., Danenov, D., & Varol, H. A. (2026) · Big Data and Cognitive Computing · doi:10.3390/bdcc10080278
Thompson, D., Mitchell, S., Adams, R., Brown, J., & Lee, M. (2026) · International Journal Of Pharmaceutical Research And Allied Sciences · doi:10.51847/p5svbmxcbk
Autonomy at War: Proportionality and the AI Race in Future Warfare
Johnson, L. L. (2026) · Journal of Military Ethics · doi:10.1080/15027570.2026.2714619 · Cited by 1
“Studies in explainable AI demonstrate that complex models, while accurate in closed environments, still possess the proclivity to misclassify ambiguous or novel data (Ji et al. 2023; Sun et al. 2024).”
Chalabi², A., Chbihi-Kaddouri, A., Chergaoui, I., Kherrab, A., Adel¹, M., Ghazi, M., & Niamane, R. (2026) · Research Square · doi:10.21203/rs.3.rs-10542333/v1
A Conversational AI Agent for Campus Information Using Retrieval Augmented Generation (RAG)
Goyal, P., Chaudhary, A., Bijalwan, V., Bharti, M., & Dwivedi, A. K. (2026) · International Conference on NextGen Data Science and Analytics (ICNDSA) · doi:10.1109/icndsa68777.2026.11652169
“…utilized as standalone conversational systems for a specific domain application, such as campus intelligence, since they rely upon pre-trained knowledge that cannot be dynamically updated and can experience hallucinations [5] when responses are generated without being grounded in factual evidence.”
Matava, C. T., Voit, B., Schmitz, A., & Engelhardt, T. (2026) · Pediatric Anesthesia · doi:10.1002/pan.70288
salazar, I., Thomas, G., Doyle, M. A., Wilde, K., Al-Shawwa, A., Chohan, S. S., Khan, W., Zahra, F., Modi, L., Sommerfeldt, M. F., et al. (2026) · doi:10.2196/preprints.108967
Zhang, Z., Zhang, Y., Lu, X., & Zhang, M. (2026) · Frontiers in Public Health · doi:10.3389/fpubh.2026.1910408 · Cited by 2
Framework for Grounding Healthcare LLMs in a Causal Knowledge Graph: A Cardiovascular ExamplePreprint
Mumtaz, U., Noor, A., & Ahmed, A. (2026) · Research Square · doi:10.21203/rs.3.rs-10720346/v1
Prehn, S., Graf, L., & Masuch, M. (2026) · Mensch und Computer (MuC) · doi:10.1145/3820253.3820254
Generative LLMs for educational recommendation
Yun, W., Bekarystankyzy, A., Kassenkhan, A. M., & Tolegenov, A. (2026) · Frontiers in Artificial Intelligence · doi:10.3389/frai.2026.1827553 · Cited by 1
“However, there is a new reliability paradox when it comes to deploying generative LLMs: models that are fluent enough to act like tutors are also likely to generate plausible but factually incorrect hallucinations (Ji et al., 2023).”
Kong, D., Funabiki, N., Kyaw, H. H. S., Kotama, I. N. D., & Zhu, Z. (2026) · Future Internet · doi:10.3390/fi18080442
“Nevertheless, neural-based systems may still produce unsupported or technically incorrect answers when the retrieved context is incomplete, ambiguous, or mis-leading [20,21].”
Han, R. (2026) · Computers in Human Behavior · doi:10.1016/j.chb.2026.109154 · Cited by 2
Fuhad, C. M., Haritha, T. H., Elappila, M., & Nagaraju, S. (2026) · SN Computer Science · doi:10.1007/s42979-026-05282-0
“● Conflicting Findings: While some studies report LLMs reduce false activations via semantic reasoning [ 12 ], others show increased hallucinations in command grounding without rule-based filters [ 83].”
Zhao, W., Chen, L., Pu, W., Liu, L., Tan, X., Li, S., Qiu, X., Fan, J., Zhu, M., & Wang, Z. (2026) · Expert Systems with Applications · doi:10.1016/j.eswa.2026.134108
A Comparative Analysis of Security Vulnerabilities and Defense Mechanisms in Large Language Models
Barek, A., Rashid, M. B., Riad, A. K. I., Francia, G., Cetinkaya, C., Shahriar, H., Khan, M. U., & Ahamed, S. I. (2026) · IEEE Annual International Computer Software and Applications Conference (COMPSAC) · doi:10.1109/compsac69091.2026.00461
“Prior research on generative model “hallucinations” has documented that even state-of-the-art models regularly produce incorrect information, exhibiting overconfidence in assertions about historical events, scientific facts, and causal relationships [30, 34].”
InfluentialTask-Specific QLoRA Adaptation in RAG-Based Information Extraction from Research Articles
Dikilitaş, Y., Sayar, A., & Göz, F. (2026) · Applied Sciences · doi:10.3390/app16168294
Designing for Critical Engagement in AI-Rich Classrooms
Nakhla, G., & Soltan, M. (2026) · Advances in computational intelligence and robotics book series · doi:10.4018/979-8-3373-7902-9.ch002
Reasoning BO: Enhancing Bayesian Optimization With the Reasoning Power of LLMs
Yang, Z., Wang, D., Ge, L., Wang, B., Fu, T., & Li, Y. (2026) · Materials Genome Engineering Advances · doi:10.1002/mgea.70086 · Cited by 15
Temporal Metamorphic Testing for RAG-Based LLMs under Evolving Knowledge
Wang, H., Towey, D., Zhang, G., Xu, J., & Adwin, J. (2026) · IEEE Annual International Computer Software and Applications Conference (COMPSAC) · doi:10.1109/compsac69091.2026.00368
“Hallucination refers to the phenomenon where language models generate incorrect or unsupported information [17].”
Al-Ansi, A. A., & Al-Wagih, K. S. (2026) · International Conference on Emerging Smart Technologies and Applications (eSmarTA) · doi:10.1109/esmarta70636.2026.11652250
“Large Language Models (LLMs) have demonstrated remarkable generative capabilities, yet they remain prone to hallucination producing fluent but factually incorrect outputs [1],[2].”
Generating knowledge graphs from news articles via multi-agent claim extraction and attribution
Fadda, M., Motta, E., Osborne, F., Recupero, D. R., & Salatino, A. A. (2026) · Knowledge-Based Systems · doi:10.1016/j.knosys.2026.116890
Tong, G.-W., Cai, J., Huang, L.-Y., Du, H., Liu, B.-H., & Xu, H.-W. (2026) · 2026 IEEE 5th International Conference on Information Technology, Big Data and Artificial Intelligence (ICIBA) · doi:10.1109/iciba68339.2026.11651821
“Foundational research in RAG has primarily focused on mitigating the hallucinations inherent in LLMs by integrating parametric memory with non-parametric external datastores [7].”
Torello, D., & Fennell, J. (2026) · ASEE Annual Conference & Exposition · doi:10.18260/1-2--59625
Assessing the Reliability of Large Language Models for Scientific Information Extraction
Schneider, L., Maitra, D., Zhao, J., & Gong, J. (2026) · ASEE Annual Conference & Exposition · doi:10.18260/1-2--58978
Teacher-in-the-Loop LLM-Enabled Instructional Scaffolding for Elementary Music Appreciation
Mao, Y., He, X., & Zheng, Y. (2026) · International Conference on Computer Science and Technologies in Education (CSTE) · doi:10.1109/cste69562.2026.11649786
“Governance measures are not optional add-ons; they are core to trustworthy classroom deployment [11], [12], [14], especially given LLM hallucination risks [15], [16].”
Halat, O. (2025) · Academic Visions · doi:10.66556/2786-586x.55.halat-o
Ajah, J., & Jamison, C. S. E. (2026) · ASEE Annual Conference & Exposition · doi:10.18260/1-2--59978
Morris, C., Chandra, K., Tripathy, S., Arias, O., & Often, E. (2026) · ASEE Annual Conference & Exposition · doi:10.18260/1-2--59142
Khan, T., & Osondu, I. S. (2026) · ASEE Annual Conference & Exposition · doi:10.18260/1-2--59130
Artificial intelligence in pharmacotherapy—Methods and approaches
Rose, O., Hoti, K., Mosch, C., & Bücker, M. (2026) · Elsevier eBooks · doi:10.1016/b978-0-443-44607-8.00030-3
Kim, R. S. (2026) · Communications AI & Computing · doi:10.1038/s44488-026-00006-y
Hu, X. (2026) · International Conference on Computer Science and Technologies in Education (CSTE) · doi:10.1109/cste69562.2026.11649770
“…the application layer exhibits insufficient domain adaptability and instruction-following bias, which frequently generates "AI illusions" — plausible yet factually inconsistent information that may contradict user inputs, thereby causing service information distortion and misdirection [15].”
Grochowalski, J. H., Ventura, C., Saraç, M., & Hendrickson, A. (2026) · Journal of Educational Measurement · doi:10.1111/jedm.70058
Hasei, J., Toda, Y., Nishimura, Y., IKEZAKI, T., Matsuoka, H., Takeuchi, K., Nakahara, R., & Ozaki, T. (2026) · Frontiers in Digital Health · doi:10.2196/preprints.109786
PRISM: stable black-box prompt optimization via multi-sampling and semantic consensus
Pham, T.-A., Cuong, N. M., Anh, N. Đ., Tran, H.-V., Nguyen, D.-A., & Hùng, P. N. (2026) · Pattern Analysis and Applications · doi:10.1007/s10044-026-01749-1
HGFuzzer: Directed Greybox Fuzzing via Large Language Model
Xu, H., Zhao, Y., Yu, J., & Wang, H. (2026) · ACM Transactions on Software Engineering and Methodology · doi:10.1145/3841476 · Cited by 18
AUDITING ARTIFICAL INTELLIGENCE
Turan, O., Müftüoğlu, Z., & Özbilge, T. (2026) · Denetişim · doi:10.58348/denetisim.1965976
Okeleye, F., Oluwaseun, B., Fagbohun, O., Sikiru, O., Akindahunsi, T., Jaiyesimi, B. S., Adetunji, R. A., Arabambi, A., & Akintola, A. S. (2026) · 2026 IEEE 9th International Conference on Big Data and Artificial Intelligence (BDAI) · doi:10.1109/bdai70753.2026.11655102
“…in financial NLP [6], [7], but applying them to high-stakes due dili-gence raises five challenges: (i) hallucination of plausible-but-false claims [8], [9]; (ii) knowledge currency , since static weights miss recent rounds and competitive shifts; (iii) multi-dimensional analysis across finance,…”
Matilla-Molina, A., Dodero, J. M., & Muñoz, A. (2026) · IEEE International Conference on Big Data Computing Service and Applications (BigDataService) · doi:10.1109/bigdataservice70481.2026.00020
“However, their reliance on static knowledge makes them vulnerable to hallucinations, outdated information, and limited factual reliability in dynamic academic settings [2].”
Improving Cost Efficiency in Medical Retrieval-Augmented Generation via Adaptive Query Routing
Chang, K., & Alireza Shafiey., M. (2026) · IEEE International Conference on Consumer Electronics - Taiwan (ICCE-Taiwan) · doi:10.1109/icce-taiwan71481.2026.11652397
Boundary-aware tool routing framework for large language models in vertical domains scenarios
Zhang, H., Shi, X., Chen, H., Zhao, X., & Ge, S. (2026) · Multimedia Systems · doi:10.1007/s00530-026-02625-y
“Particularly when user instructions are semantically ambiguous, models are prone to problems such as incorrect tool invocation, excessive invocation, or non-standard parameter generation [8, 9].”
From perception to governance: Knowledge-enhanced intelligence for smart sustainable buildings
Liu, S., Xie, H., Lei, Y., Zhao, J., Liu, J., Song, G., Zhang, L., Yang, W., Song, X., & Zhang, H. (2026) · Energy and Buildings · doi:10.1016/j.enbuild.2026.118128
AI Hallucinations in Academic Writing and Addresses
Liu, Y. (2026) · SHS Web of Conferences · doi:10.1051/shsconf/202623602006
An Entailment-Graph-Based Method for Black-Box Uncertainty Estimation in Large Language Models
Kong, Y., Yang, X., Ma, X., & Wang, M. (2026) · IEEE International Conference on Big Data and Artificial Intelligence (BDAI) · doi:10.1109/bdai70753.2026.11655168
“Large language models (LLMs) have become foundation components for reasoning, question answering, and decision support, yet their fluent outputs can still contain factual fabrications or unsupported conclusions [1] – [4].”
SAS Multimetric Tribunal: Structural Classification of Hallucination in AI-Generated TextPreprint
Durante, G. E. (2026) · Qeios · doi:10.32388/t420iv.2
The Semantic Power Factor (SPF): A Complex AC Power Model for Generative AI Communication
Tantisukarom, C. (2026) · Systems Research and Behavioral Science · doi:10.1002/sres.70149
TLC-RAG: A Plain Text RAG Method Based on Triples and Logic Chains
Yang, W., Cui, L., & Xue, N. (2026) · IEEE International Conference on Big Data and Artificial Intelligence (BDAI) · doi:10.1109/bdai70753.2026.11655161
“However, it has three core limitations for high-precision scenarios: low info utilization (redundant text raises computing cost/omits key info [3]), weak reasoning (unstructured chunks hinder knowledge integration [4]), and high hallucination risk (LLMs fabricate facts with vague info [5]).”
Oh, H., Salcedo-Vázquez, A. M., Ramírez-Alpizar, I. G., & Domae, Y. (2026) · IEEE Access · doi:10.1109/access.2026.3725998
“Hallucina-tions, which manifest as inconsistent outlier outputs, are a key issue in LLMs [6], and the production of hallucinated trajectories is particularly problematic for the reliability of robotic applications.”
Tanakov, I., Zoubovsky, E., Berry, I., & Woelfel-Monsivais, D. (2026) · SPE Artificial Lift Conference and Exhibition - Americas (ALCE) · doi:10.2118/232983-ms
Fakher, L., & Bauer, A. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-37654-1_31
Srivastava, P. (2026) · IEEE International Conference on Big Data and Artificial Intelligence (BDAI) · doi:10.1109/bdai70753.2026.11655132
“Ji and colleagues provide a comprehensive taxonomy of hallucination types in natural language generation, noting that factual inconsistency is most prevalent in summarization and document drafting tasks [6].”
Moral, A., Antonio Arroyo, S., Aparicio, J., & Barber, X. (2026) · Machine Learning and Knowledge Extraction · doi:10.3390/make8090256
Kamińska, A. M. (2026) · Future Collections Libraries Archives and Museums · doi:10.3390/fclam1010003
Review of Large Language Model-Enhanced Curriculum Design in E-Learning
Simon, B., Hartvég, Á., Szász, L., Dénes-Fazakas, L., Eigner, G., & Kovács, L. (2026) · International Conference on Intelligent Engineering Systems · doi:10.1109/ines69513.2026.11661260
“…been shown to produce content with higher rates of factual inaccuracies, weaker alignment with specific learning objectives, and reduced pedagogical appropriateness, as the absence of expert judgment allows hallucinations and generic outputs to propagate unchecked into final materials [4], [20].”
Ocak, A., Diri, S., & Kilimci, Z. H. (2026) · Süleyman Demirel Üniversitesi Fen Bilimleri Enstitüsü Dergisi · doi:10.19113/sdufenbed.1912302
Sanjalawe, Y. K., Al-E’mari, S., & Makhadmeh, S. N. (2026) · Evolutionary Intelligence · doi:10.1007/s12065-026-01255-7
“Ji et al. [12] provide the canonical taxonomy, distinguishing between intrinsic hallucination (output contradicts the source context) and extrinsic hallucination (output introduces facts absent from context).”
InfluentialLiu, Y., Zhu, L., Deng, H., Chen, Q., Chen, X., Zhang, S., & Song, C. (2026) · Buildings · doi:10.3390/buildings16173391
Christlin Jefrina, R., Mythily, M., M, V. I., & Iwin Thanakumar Joseph, S. (2026) · Journal of Advances in Information Technology · doi:10.12720/jait.17.8.1530-1543
Causal attributions for ChatGPT errors: what is the role of AI literacy?
Norcia, M., & Colí, E. (2026) · Journal of Psychology and AI · doi:10.1080/29974100.2026.2712302
Faithful open-domain QA via prompt-based multi-source knowledge fusion
Zhang, Y., Chen, Y., Zheng, D., Tu, Y., & Wenju, Q. (2026) · Journal of Intelligent Information Systems · doi:10.1007/s10844-026-01084-1
Szappan, A., & Krutilla, Z. (2026) · IEEE International Conference on Intelligent Engineering Systems (INES) · doi:10.1109/ines69513.2026.11661120
“Its primary advantage lies in its simplicity of implementation; however, it may result in higher response variability and an increased risk of hallucination [13].”
Deng, Y. (2026) · 2026 3rd World Conference on Computer and Information Security (WCCIS) · doi:10.1109/wccis70285.2026.11651016
“Simultaneously, restricted by the timeliness and credibility of training data and paradigms, LLMs are prone to "hallucinations"—generating content inconsistent with facts [4], [5].”
Kapizov, D., & Kumar, P. (2026) · Scientific Journal of Astana IT University · doi:10.37943/xsaj1029
Mao, X., Wang, J., Wang, Z., Zhang, Y., Qiu, S., Ye, J., He, S., Wang, C., Xia, Y., He, C., et al. (2026) · Journal of Medical Internet Research · doi:10.2196/96002
“Even with workbook support, AI agents occasionally generated unsupported section references, overinterpreted ambiguous text, and were sensitive to the location of relevant information in long documents, consistent with the “lost in the middle” phenomenon [55,56,58].”
Nguyen, P., Wu, Y., Anderson, B., Mayorga‐Carlin, M., Sorkin, J., Yesha, Y., & Lal, B. (2026) · IEEE International Conference on Digital Health (ICDH) · doi:10.1109/icdh72779.2026.00010
“Large language models (LLMs) can perform strong medical-text reasoning, but direct LLM-as-classifier approaches are sensitive to prompt phrasing, hallucination, and may lack calibrated uncertainty [3], [4].”
Κουτσουπιάς, Ν., & Nosios, M. (2026) · Advances in computational intelligence and robotics book series · doi:10.4018/979-8-3373-7262-4.ch001
Durán, A. G., de Oliver-Lopez, C. P., & Bravo, C. B. (2026) · Journal of Cultural Heritage · doi:10.1016/j.culher.2026.08.010 · Cited by 1
Wang, Z., Zhang, H., Hou, J., & Zheng, S. (2026) · International Conference on Electronics and Communication, Network and Computer Technology (ECNCT) · doi:10.1109/ecnct70535.2026.11661501
LLMs as an Abstraction Layer in Computer Organization: Implementation and Limitations
Hoffmann, H., & Spesse, C. M. P. (2026) · doi:10.5753/semish.2026.21960
“Authors such as [Ji et al. 2023] and [Maynez et al. 2020] propose a taxonomy in which hallucinations can be classified into two categories: intrinsic and extrinsic.”
Alucinações em Modelos de Linguagem de Grande Escala: Uma Revisão Sistemática da Literatura
Pereira, A. S., Aquino, S. A. B. D. M., & Xavier, E. S. (2026) · Anais do LIII Seminário Integrado de Software e Hardware (SEMISH 2026) · doi:10.5753/semish.2026.23904
“As categorias semântica e linguística, embora definidas em surveys como Ji et al. (2023) e Huang et al. (2023), carecem de metodologias experimentais dedicadas.”
InfluentialQian, X., Xu, Y., Ma, C., & Chan, C. K. (2026) · Research Square · doi:10.21203/rs.3.rs-10522439/v1
Self-citationShih, P. F. (2026) · Japanese Journal of Radiology · doi:10.1007/s11604-026-02059-6 · Cited by 2
The TRACE Model for Graduate Research Ethics, Publication Integrity, and Global Academic Contexts
KALECİ, F. (2026) · Advances in computational intelligence and robotics book series · doi:10.4018/979-8-2600-1232-1.ch004 · Cited by 1
The Pyramid Within: How Whole-Person Leaders Endure in an AI-Accelerated Future
Migliore, L. (2026) · Advances in computational intelligence and robotics book series · doi:10.4018/979-8-3373-7709-4.ch002
Industry expectations for entry-level AI skills: evidence for business curriculum designPreprint
Abbasi-Pooya, A., Fish, K. E., & Ruby, P. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7357801
Anderson, M., Wong, L., & Lee, S. (2025) · Interdisciplinary Research in Medical Sciences Specialty · doi:10.51847/o8uzmkyrdq
Li, M., Wang, H., Sun, Z., & Zhou, H. (2026) · Big Data and Cognitive Computing · doi:10.3390/bdcc10090291
Zhorzholiani, T. (2026) · Advances in computational intelligence and robotics book series · doi:10.4018/979-8-2600-1232-1.ch009
Explaining Driver Behavior in Sim Racing with Shannon Entropy and LLM Feedback
Nunes, T., Medeiros, M. D. A., Silva, M., Bittencourt, J. C. N., Costa, D. G., & da Silva, I. M. D. (2026) · Entropy · doi:10.3390/e28090960
“In addition, when LLMs are used, the boundary between analytical inference and semantic communication is not always clear, increasing the risk of domain-specific hallucinations [17].”
LLM-Based Tool for Automatic Feedback for Rubric-Based Feedback in Construction
Yang, Z., Castelblanco, G., & Cruz-Castro, L. (2026) · ASCE Construction Research Congress (CRC) · doi:10.1061/9780784486962.033
Metadata
- Authors
- Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P.
- Affiliations
- Hong Kong University of Science and Technology (HK)
- ORCID
- Venue
- ACM Computing Surveys
- Published
- 2023-03-03
- Volume
- 55
- Issue
- 12
- Pages
- 1-38
- ISSN
- 0360-0300 (ISSN-L)
- 1557-7341
- OpenAlex
- W4309674289
- Semantic Scholar
- 3def68bd0f856886d34272840a7f81588f2bc082
- Cited by
- 5,025
- References
- 111
- Citations by year
- 2026: 2296 · 2025: 1441 · 2024: 817 · 2023: 453
- FWCI
- 224.36
- Type
- journal article
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- ACM Computing Surveys (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- Rare & Special e-Zone (The Hong Kong University of Science and Technology) (submittedVersion)
- arXiv (Cornell University) (cc-by)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Physical Sciences › Computer Science › Artificial Intelligence › Speech and dialogue systems
- Keywords
- hallucinations, natural language generation, hallucination mitigation, abstractive summarization, dialogue generation, generative question answering, data-to-text generation, machine translation, large language models, transformer models
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-09-30, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-09-30 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.




