A Benchmarking Framework for Detecting and Reducing Hallucinations in Large Language Models

This paper has proposed a unified benchmarking framework for detecting and reducing hallucinations in Large Language Models, addressing the fragmentation of current evaluation practice in which benchmarks, detectors, and mitigation techniques are studied in isolation. The framework specifies a five-category hallucination taxonomy, a multi-task benchmark composition, a claim-level annotation pipeline, and a common evaluation protocol, and conceptually compares six detection approaches — token probability and entropy scoring, sampling-based self-consistency, semantic entropy, NLI verification, retrieval-augmented fact-checking, and LLM-as-a-judge — together with five mitigation strategies. Because no models were executed and no annotated data was collected for this study, the work is presented explicitly as a proposed framework and experimental design rather than a report of measured results, and all illustrative figures and tables are labelled accordingly. If implemented, the framework is expected to give model developers, deployers, and regulators a consistent, reproducible way to measure how often LLMs hallucinate and which techniques most effectively reduce it. Realizing this potential will require addressing the limitations discussed in Section X, particularly annotation cost, knowledge coverage, and contamination, and pursuing the future directions outlined in Section XI, from multilingual and multimodal extension to an open public leaderboard.

[1]  Z. Ji et al., “Survey of hallucination in natural language generation,” ACM Computing Surveys, vol. 55, no. 12, Art. 248, 2023, doi: 10.1145/3571730. [2] L. Huang et al., “A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,” ACM Transactions on Information Systems, vol. 43, no. 2, 2025, doi: 10.1145/3703155. [3] Y. Zhang et al., “Siren's song in the AI ocean: A survey on hallucination in large language models,” arXiv:2309.01219, 2023. [4]  S. M. T. I. Tonmoy et al., “A comprehensive survey of hallucination mitigation techniques in large language models,” arXiv:2401.01313, 2024. [5] J. Maynez, S. Narayan, B. Bohnet, and R. McDonald, “On faithfulness and factuality in abstractive summarization,” in Proc. ACL, 2020, pp. 1906–1919, doi: 10.18653/v1/2020.acl-main.173. [6] S. Lin, J. Hilton, and O. Evans, “TruthfulQA: Measuring how models mimic human falsehoods,” in Proc. ACL, 2022, pp. 3214–3252, doi: 10.18653/v1/2022.acl-long.229. [7] 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. EMNLP, 2023, pp. 6449–6464, doi: 10.18653/v1/2023.emnlp-main.397. [8] J. Thorne, A. Vlachos, C. Christodoulopoulos, and A. Mittal, “FEVER: A large-scale dataset for fact extraction and VERification,” in Proc. NAACL-HLT, 2018, pp. 809–819, doi: 10.18653/v1/N18-1074. [9]  S. Min et al., “FActScore: Fine-grained atomic evaluation of factual precision in long form text generation,” in Proc. EMNLP, 2023, pp. 12076–12100, doi: 10.18653/v1/2023.emnlp-main.741. [10] J. Wei et al., “Long-form factuality in large language models,” in Proc. NeurIPS, 2024, arXiv:2403.18802.