CoChat
Sign Up
PreprintCited by 129Topic ModelingNatural Language Processing TechniquesMultimodal Machine Learning Applications

A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models

Open access
Tonmoy, S. M Towhidul Islam, Zaman, S M Mehedi, Jain, Vinija, Rani, Anku, Rawte, Vipula, Chadha, Aman, Das, Amitava·2024·arXiv
DOI 10.48550/arxiv.2401.01313arXiv 2401.01313
Open this paper in CoChat — free accountOpen PDF

Abstract

As Large Language Models (LLMs) continue to advance in their ability to write human-like text, a key challenge remains around their tendency to hallucinate generating content that appears factual but is ungrounded. This issue of hallucination is arguably the biggest hindrance to safely deploying these powerful LLMs into real-world production systems that impact people's lives. The journey toward widespread adoption of LLMs in practical settings heavily relies on addressing and mitigating hallucinations. Unlike traditional AI systems focused on limited tasks, LLMs have been exposed to vast amounts of online text data during training. While this allows them to display impressive language fluency, it also means they are capable of extrapolating information from the biases in training data, misinterpreting ambiguous prompts, or modifying the information to align superficially with the input. This becomes hugely alarming when we rely on language generation capabilities for sensitive applications, such as summarizing medical records, financial analysis reports, etc. This paper presents a comprehensive survey of over 32 techniques developed to mitigate hallucination in LLMs. Notable among these are Retrieval Augmented Generation (Lewis et al, 2021), Knowledge Retrieval (Varshney et al,2023), CoNLI (Lei et al, 2023), and CoVe (Dhuliawala et al, 2023). Furthermore, we introduce a detailed taxonomy categorizing these methods based on various parameters, such as dataset utilization, common tasks, feedback mechanisms, and retriever types. This classification helps distinguish the diverse approaches specifically designed to tackle hallucination issues in LLMs. Additionally, we analyze the challenges and limitations inherent in these techniques, providing a solid foundation for future research in addressing hallucinations and related phenomena within the realm of LLMs.

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

DataCite · 10.48550/arxiv.2401.01313

Cites

Statistics

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

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?

  • LLMs have well-documented prompt sensitivity, where even minor variations in phrasing can lead to divergent behaviors.

    “Compounding this is the well-documented prompt sensitivity of LLMs, where even minor variations in phrasing can lead to divergent behaviors.”

    From AI Agents vs. Agentic AI: A Conceptual taxonomy, applications and challenges (2026) · cites this paper as [192]

    Section 2.1, “Prompt Engineering” (page unavailable), describes experimenting with different instructions to improve model output, supporting the general point that prompts affect responses. It does not say that minor changes in phrasing can produce divergent behaviors. (LLM, conf=0.81)

    Experimental — accuracy not yet measured

  • This brittleness hampers reproducibility, necessitating meticulous manual prompt engineering and often requiring domain-specific tuning to maintain consistency across interactions.

    “This brittleness hampers reproducibility, necessitating meticulous manual prompt engineering and often requiring domain-specific tuning to maintain consistency across interactions [192].”

    From AI Agents vs. Agentic AI: A Conceptual taxonomy, applications and challenges (2026) · cites this paper as [192]

    Section 2.1 describes prompt engineering as experimenting with instructions to improve outputs, and §2.1.3 describes tuning prompts for specific tasks. The copy does not establish that brittleness hampers reproducibility or that meticulous manual engineering and domain-specific tuning are often necessary to maintain consistency across interactions. No page numbers are available in the indexed passages. (LLM, conf=0.78)

    Experimental — accuracy not yet measured

  • Such errors undermine scholarly trust and threaten the integrity of scientific communication.

    “Such errors undermine scholarly trust and threaten the integrity of scientific communication [8, 25].”

    From CiteAudit: You Cited It, But Did You Read It? A Benchmark for Verifying Scientific References in the LLM Era (2026) · cites this paper as [25]

    The Introduction describes LLM use in academic research and warns that hallucinations can produce inaccurate or wholly unfounded references. Section 2.1.1.3 also links unsupported content to difficulty judging LLM outputs’ trustworthiness. These passages support a concern about trust, but the paper does not specifically discuss the integrity of scientific communication. (LLM, conf=0.78)

    Experimental — accuracy not yet measured

  • RAG reduces the frequency of hallucinations and improves interpretability, especially for information-sensitive tasks such as question answering and multi-hop reasoning.

    “By combining both parametric (trained) and non-parametric (retrieved) sources, RAG reduces the frequency of hallucinations and improves interpretability, especially for information-sensitive tasks such as question answering and multi-hop reasoning [4, 25, 33].”

    From SKiM-GPT: combining biomedical literature-based discovery with large language model hypothesis evaluation (2025) · cites this paper as [33]

    Section 2.1.1 says RAG uses external knowledge and mitigates hallucinations, while §2.1.1.4 describes combining a trained generator with retrieved documents for knowledge-intensive tasks. The survey discusses question answering and multi-hop questions among retrieval-based techniques (§5), but it does not establish that RAG improves interpretability; it says RAG's responses are verifiable, which is a different claim. No page numbers are available in our copy. (LLM, conf=0.83)

    Experimental — accuracy not yet measured

  • Large language models hallucinate factual content even when surface fluency is maintained.

    “Large language models hallucinate factual content even when surface fluency is maintained, a failure mode characterized across model families, training regimes, and deployment settings in recent surveys [14, 29, 19] and in zero-resource detection work such as SelfCheckGPT [17].”

    From Source or It Didn't Happen: A Multi-Agent Framework for Citation Hallucination Detection (2026) · cites this paper as [29]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Cites

No references could be resolved yet.

Research works printed but not linked: 17

  1. Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. 2023. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. arXiv preprint arXiv:2311.10081. · arXiv
  2. George Chrysostomou and Nikolaos Aletras. 2021. En- joy the salience: Towards better transformer-based faithful explanations with word salience. arXiv preprint arXiv:2108.13759. · arXiv
  3. Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan, Vincent Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, et al. 2023. Rarr: Researching and revising what language models say, using language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16477-16508.
  4. Ziwei Ji, Zihan Liu, Nayeon Lee, Tiezheng Yu, Bryan Wilie, Min Zeng, and Pascale Fung. 2023a. RHO: Reducing hallucination in open-domain dialogues with knowledge grounding. In Findings of the As- sociation for Computational Linguistics: ACL 2023, pages 4504-4522, Toronto, Canada. Association for Computational Linguistics. · DOI · Paper page
  5. Mateusz Lango and Ondrej Dusek. 2023. Critic-driven decoding for mitigating hallucinations in data-to-text generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Process- ing, pages 2853-2862, Singapore. Association for Computational Linguistics. · DOI
  6. Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045-3059, Online and Punta Cana, Domini- can Republic. Association for Computational Lin- guistics. · DOI · Paper page
  7. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2021. Retrieval-augmented generation for knowledge-intensive nlp tasks. arXiv preprint arXiv:2005.11401v4. · arXiv
  8. Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023a. Inference- time intervention: Eliciting truthful answers from a language model. arXiv preprint arXiv:2306.03341. · arXiv
  9. Niels Mündler, Jingxuan He, Slobodan Jenko, and Mar- tin Vechev. 2023. Self-contradictory hallucinations of large language models: Evaluation, detection and mitigation. arXiv preprint arXiv:2305.15852. · arXiv
  10. Yifu Qiu, Varun Embar, Shay B Cohen, and Benjamin Han. 2023a. Think while you write: Hypothesis verification promotes faithful knowledge-to-text gen- eration. arXiv preprint arXiv:2311.09467. · arXiv
  11. Partha Pratim Ray. 2023. Chatgpt: A comprehensive review on background, applications, key challenges, bias, ethics, limitations and future scope. Internet of Things and Cyber-Physical Systems, 3:121-154. · DOI · Paper page
  12. Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108. · arXiv
  13. Chenglei Si, Zhe Gan, Zhengyuan Yang, Shuohang Wang, Jianfeng Wang, Jordan Boyd-Graber, and Li- juan Wang. 2022. Prompting gpt-3 to be reliable. arXiv preprint arXiv:2210.09150. · arXiv
  14. Kai Sun, Yifan Ethan Xu, Hanwen Zha, Yue Liu, and Xin Luna Dong. 2023a. Head-to-tail: How knowl- edgeable are large language models (llm)? aka will llms replace knowledge graphs? arXiv preprint arXiv:2308.10168. · arXiv
  15. BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilić, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luc- cioni, François Yvon, et al. 2022. Bloom: A 176b- parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100. · arXiv
  16. Sunjae Yoon, Eunseop Yoon, Hee Suk Yoon, Junyeong Kim, and Chang Yoo. 2022. Information-theoretic text hallucination reduction for video-grounded di- alogue. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4182-4193, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. · DOI
  17. Hanning Zhang, Shizhe Diao, Yong Lin, Yi R Fung, Qing Lian, Xingyao Wang, Yangyi Chen, Heng Ji, and Tong Zhang. 2023a. R-tuning: Teaching large language models to refuse unknown questions. arXiv preprint arXiv:2311.09677. · arXiv

Cited by

Sorted by citations

Authors
Tonmoy, S. M. T. I., Zaman, S. M. M., Jain, V., Rani, A., Rawte, V., Chadha, A., & Das, A.
Affiliations
  • Islamic University of Technology (BD)
  • Amazon (United States) (US)
  • University of South Carolina (US)
ORCID
Venue
arXiv
Publisher
arXiv
Published
2024
ISSN
2331-8422 (ISSN-L)
DOI
10.48550/arxiv.2401.01313
arXiv
2401.01313
OpenAlex
W4390602555
Cited by
129
References
0
Citations by year
2026: 26 · 2025: 57 · 2024: 45 · 2021: 1
Type
posted content
Language
en
Open access
Licence
Locations
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • 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 › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
Keywords
hallucination mitigation, large language models, retrieval-augmented generation, knowledge retrieval, hallucination taxonomy, training data bias, factual grounding

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.

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