Browse All Articles
Review Article

Fluent but Unfounded: A Review of Hallucination in Small Language Models

*1ORCID

* Corresponding author

*1 Independent Researcher [email protected]
Small language models Hallucination Factuality Retrieval-augmented generation Uncertainty estimation Abstention AI governance Misinformation

Abstract

Small language models (SLMs), compact generative models ranging from a few hundred million to roughly ten billion parameters, are increasingly chosen for on-device, private and cost-sensitive deployments. Their tendency to produce fluent but unsupported statements is nevertheless rarely treated as a distinct engineering problem. This review synthesises research published up to November 2026 on why SLMs hallucinate, how the problem is measured, and which countermeasures exist at the model, inference and operational levels. Three arguments are made. First, the evidence points to several interacting causes, namely limited capacity for rare facts, training and evaluation incentives that reward guessing, unstable retrieval of knowledge the model does hold, and weak calibration, rather than to a single deficit. Second, reported hallucination rates depend strongly on model variant, decoding settings and task, so figures cannot be compared across studies without care. Third, no single safeguard suffices: reliability is better pursued as a layered stack that combines targeted fine-tuning, retrieval grounding, sampling-based consistency checks, abstention, and risk-proportionate human oversight. The review then maps these layers onto accountability and regulatory expectations and lists open problems, including size-controlled benchmarks and independent replication of promising small-scale methods.

References

  1. Z. Lu et al., "Small language models: Survey, measurements, and insights," arXiv:2409.15790, 2024. https://doi.org/10.48550/arXiv.2409.15790
  2. F. Wang et al., "A comprehensive survey of small language models in the era of large language models: Techniques, enhancements, applications, collaboration with LLMs, and trustworthiness," ACM Trans. Intell. Syst. Technol., vol. 16, no. 6, pp. 1–87, 2025.
  3. P. Belcak et al., "Small language models are the future of agentic AI," arXiv:2506.02153, 2025. https://doi.org/10.48550/arXiv.2506.02153
  4. M. Abdin et al., "Phi-3 technical report: A highly capable language model locally on your phone," arXiv:2404.14219, 2024. https://doi.org/10.48550/arXiv.2404.14219
  5. Gemma Team, "Gemma 2: Improving open language models at a practical size," arXiv:2408.00118, 2024. https://doi.org/10.48550/arXiv.2408.00118
  6. A. Yang et al., "Qwen3 technical report," arXiv:2505.09388, 2025. https://doi.org/10.48550/arXiv.2505.09388
  7. A. Q. Jiang et al., "Mistral 7B," arXiv:2310.06825, 2023. https://doi.org/10.48550/arXiv.2310.06825
  8. P. Zhang, G. Zeng, T. Wang, and W. Lu, "TinyLlama: An open-source small language model," arXiv:2401.02385, 2024. https://doi.org/10.48550/arXiv.2401.02385
  9. Z. Ji et al., "Survey of hallucination in natural language generation," ACM Comput. Surv., vol. 55, no. 12, Art. no. 248, pp. 1–38, 2023. https://doi.org/10.1145/3571730
  10. L. Huang et al., "A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions," ACM Trans. Inf. Syst., vol. 43, no. 2, Art. no. 42, 2025. https://doi.org/10.1145/3703155
  11. Y. Zhang et al., "Siren's song in the AI ocean: A survey on hallucination in large language models," arXiv:2309.01219, 2023. https://doi.org/10.48550/arXiv.2309.01219
  12. E. Ricco, E. Onofri, L. Cima, S. Cresci, and R. Di Pietro, "A geometric analysis of small-sized language model hallucinations," arXiv:2602.14778, 2026; presented at Int. Conf. Mach. Learn. (ICML), 2026. https://doi.org/10.48550/arXiv.2602.14778
  13. A. V. Aravindan and M. Kejriwal, "Second Guess: Detecting uncertainty through abstention and answer stability in small language models," arXiv:2605.25394, 2026. https://doi.org/10.48550/arXiv.2605.25394
  14. M. A. Baker and F. Al-Qrize, "Mitigating hallucination in small language models via contrastive chain-of-thought fine-tuning," SOLAV J., vol. 1, pp. 33–53, Jul. 2025. https://doi.org/10.5281/zenodo.18538736
  15. M. A. Baker, "The spread of AI-generated misinformation," Authorea, preprint, Apr. 7, 2025. https://doi.org/10.22541/au.174405164.43017701/v1
  16. C. Drolsbach and N. Pröllochs, "Characterizing AI-generated misinformation on social media," arXiv:2505.10266, 2025. https://doi.org/10.48550/arXiv.2505.10266
  17. Moffatt v. Air Canada, 2024 BCCRT 149 (Civil Resolution Tribunal of British Columbia, Feb. 14, 2024). https://canlii.ca/t/k2b02
  18. W. H. Walters and E. I. Wilder, "Fabrication and errors in the bibliographic citations generated by ChatGPT," Sci. Rep., vol. 13, Art. no. 14045, 2023. https://doi.org/10.1038/s41598-023-41032-5
  19. M. Dahl, V. Magesh, M. Suzgun, and D. E. Ho, "Large legal fictions: Profiling legal hallucinations in large language models," J. Legal Anal., vol. 16, no. 1, pp. 64–93, 2024.
  20. V. Magesh, F. Surani, M. Dahl, M. Suzgun, C. D. Manning, and D. E. Ho, "Hallucination-free? Assessing the reliability of leading AI legal research tools," J. Empir. Leg. Stud., vol. 22, no. 2, pp. 216–242, 2025.
  21. M. A. Baker, "AI ethics in historical research: A framework for bias mitigation, transparency, and accountability," SSRN, Oct. 5, 2025. https://doi.org/10.2139/ssrn.5565499
  22. S. Farquhar, J. Kossen, L. Kuhn, and Y. Gal, "Detecting hallucinations in large language models using semantic entropy," Nature, vol. 630, no. 8017, pp. 625–630, 2024. https://doi.org/10.1038/s41586-024-07421-0
  23. S. Lin, J. Hilton, and O. Evans, "TruthfulQA: Measuring how models mimic human falsehoods," in Proc. 60th Annu. Meeting Assoc. Comput. Linguistics (ACL), 2022, pp. 3214–3252. https://doi.org/10.18653/v1/2022.acl-long.229
  24. J. Li, X. Cheng, W. X. Zhao, J.-Y. Nie, and J.-R. Wen, "HaluEval: A large-scale hallucination evaluation benchmark for large language models," in Proc. Conf. Empir. Methods Natural Lang. Process. (EMNLP), 2023, pp. 6449–6464. https://doi.org/10.18653/v1/2023.emnlp-main.397
  25. S. Min et al., "FActScore: Fine-grained atomic evaluation of factual precision in long form text generation," in Proc. EMNLP, 2023, pp. 12076–12100. https://doi.org/10.18653/v1/2023.emnlp-main.741
  26. P. Manakul, A. Liusie, and M. Gales, "SelfCheckGPT: Zero-resource black-box hallucination detection for generative large language models," in Proc. EMNLP, 2023, pp. 9004–9017. https://doi.org/10.18653/v1/2023.emnlp-main.557
  27. Y. Bang et al., "HalluLens: LLM hallucination benchmark," in Proc. 63rd Annu. Meeting ACL, 2025, pp. 24128–24156.
  28. Z. Allen-Zhu and Y. Li, "Physics of language models: Part 3.3, knowledge capacity scaling laws," arXiv:2404.05405, 2024. https://doi.org/10.48550/arXiv.2404.05405
  29. N. Kandpal, H. Deng, A. Roberts, E. Wallace, and C. Raffel, "Large language models struggle to learn long-tail knowledge," in Proc. 40th Int. Conf. Mach. Learn. (ICML), 2023, pp. 15696–15707. https://proceedings.mlr.press/v202/kandpal23a.html
  30. A. Mallen et al., "When not to trust language models: Investigating effectiveness of parametric and non-parametric memories," in Proc. 61st Annu. Meeting ACL, 2023, pp. 9802–9822. https://doi.org/10.18653/v1/2023.acl-long.546
  31. A. T. Kalai, O. Nachum, S. S. Vempala, and E. Zhang, "Why language models hallucinate," arXiv:2509.04664, 2025. https://doi.org/10.48550/arXiv.2509.04664
  32. A. T. Kalai and S. S. Vempala, "Calibrated language models must hallucinate," in Proc. 56th Annu. ACM Symp. Theory Comput. (STOC), 2024, pp. 160–171.
  33. Z. Xu, S. Jain, and M. Kankanhalli, "Hallucination is inevitable: An innate limitation of large language models," arXiv:2401.11817, 2024. https://doi.org/10.48550/arXiv.2401.11817
  34. S. Kadavath et al., "Language models (mostly) know what they know," arXiv:2207.05221, 2022. https://doi.org/10.48550/arXiv.2207.05221
  35. S. Lin, J. Hilton, and O. Evans, "Teaching models to express their uncertainty in words," Trans. Mach. Learn. Res., 2022. https://openreview.net/forum?id=8s8K2UZGTZ
  36. M. Xiong et al., "Can LLMs express their uncertainty? An empirical evaluation of confidence elicitation in LLMs," in Proc. Int. Conf. Learn. Represent. (ICLR), 2024. https://openreview.net/forum?id=gjeQKFxFpZ
  37. Y. Elazar et al., "Measuring and improving consistency in pretrained language models," Trans. Assoc. Comput. Linguistics, vol. 9, pp. 1012–1031, 2021. https://doi.org/10.1162/tacl_a_00410
  38. M. Sclar, Y. Choi, Y. Tsvetkov, and A. Suhr, "Quantifying language models' sensitivity to spurious features in prompt design or: How I learned to start worrying about prompt formatting," in Proc. ICLR, 2024. https://openreview.net/forum?id=RIu5lyNXjT
  39. N. F. Liu et al., "Lost in the middle: How language models use long contexts," Trans. Assoc. Comput. Linguistics, vol. 12, pp. 157–173, 2024. https://doi.org/10.1162/tacl_a_00638
  40. Z. Gekhman et al., "Does fine-tuning LLMs on new knowledge encourage hallucinations?" in Proc. EMNLP, 2024. https://doi.org/10.48550/arXiv.2405.05904
  41. L. Ouyang et al., "Training language models to follow instructions with human feedback," in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 35, 2022, pp. 27730–27744. https://doi.org/10.48550/arXiv.2203.02155
  42. J. Wei et al., "Chain-of-thought prompting elicits reasoning in large language models," in Proc. NeurIPS, vol. 35, 2022, pp. 24824–24837. https://doi.org/10.48550/arXiv.2201.11903
  43. S. Gunasekar et al., "Textbooks are all you need," arXiv:2306.11644, 2023. https://doi.org/10.48550/arXiv.2306.11644
  44. E. J. Hu et al., "LoRA: Low-rank adaptation of large language models," in Proc. ICLR, 2022. https://openreview.net/forum?id=nZeVKeeFYf9
  45. T. Dettmers, A. Pagnoni, A. Holtzman, and L. Zettlemoyer, "QLoRA: Efficient finetuning of quantized LLMs," in Proc. NeurIPS, vol. 36, 2023. https://doi.org/10.48550/arXiv.2305.14314
  46. R. Rafailov et al., "Direct preference optimization: Your language model is secretly a reward model," in Proc. NeurIPS, vol. 36, 2023. https://doi.org/10.48550/arXiv.2305.18290
  47. C. Tomani, K. Chaudhuri, I. Evtimov, D. Cremers, and M. Ibrahim, "Uncertainty-based abstention in LLMs improves safety and reduces hallucinations," arXiv:2404.10960, 2024. https://doi.org/10.48550/arXiv.2404.10960
  48. B. Wen et al., "Know your limits: A survey of abstention in large language models," Trans. Assoc. Comput. Linguistics, vol. 13, pp. 529–556, 2025.
  49. S. M. T. I. Tonmoy et al., "A comprehensive survey of hallucination mitigation techniques in large language models," arXiv:2401.01313, 2024. https://doi.org/10.48550/arXiv.2401.01313
  50. P. Lewis et al., "Retrieval-augmented generation for knowledge-intensive NLP tasks," in Proc. NeurIPS, vol. 33, 2020, pp. 9459–9474. https://doi.org/10.48550/arXiv.2005.11401
  51. Y. Gao et al., "Retrieval-augmented generation for large language models: A survey," arXiv:2312.10997, 2023. https://doi.org/10.48550/arXiv.2312.10997
  52. X. Wang et al., "Self-consistency improves chain of thought reasoning in language models," in Proc. ICLR, 2023. https://doi.org/10.48550/arXiv.2203.11171
  53. L. Kuhn, Y. Gal, and S. Farquhar, "Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation," in Proc. ICLR, 2023. https://openreview.net/forum?id=VD-AYtP0dve
  54. S. Dhuliawala et al., "Chain-of-verification reduces hallucination in large language models," arXiv:2309.11495, 2023. https://doi.org/10.48550/arXiv.2309.11495
  55. National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile," NIST AI 600-1, Jul. 2024. https://doi.org/10.6028/NIST.AI.600-1
  56. Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Off. J. Eur. Union, Jul. 12, 2024. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
  57. E. M. Bender, T. Gebru, A. McMillan-Major, and S. Shmitchell, "On the dangers of stochastic parrots: Can language models be too big?," in Proc. ACM Conf. Fairness, Accountability, Transparency (FAccT), 2021, pp. 610–623. https://doi.org/10.1145/3442188.3445922
Download PDF This DOI is provided for long-term preservation and open archiving. (Zenodo) Archival DOI (Zenodo)
Share This Article
Article Info
  • Received: 2026-10-07
  • Accepted: 2026-11-09
  • Published: 2026-11-12
  • Pages: 170-185
  • Citations: 1
  • Type: Review Article
  • Volume: 2
  • Version: 2026-11-12 (1)
  • License: CC BY 4.0