大语言模型幻觉检测方法综述

Summary

This survey, from the National University of Defense Technology, looks at how to detect hallucinations in large language models (LLMs). Hallucinations are outputs that read fluently but are unfaithful, false or self-contradictory. They are a serious risk in sensitive domains such as law, finance and the military, and they are harder to catch than errors in older NLP systems. The authors first give a sharper definition and a finer taxonomy. They then trace where hallucinations come from across the whole model lifecycle: architecture, pretraining, fine-tuning, alignment and inference. Detection methods are grouped by how much access one has to the model, which gives white-box approaches (using internal states) and black-box approaches (using outputs only). The survey compares their strengths and limits for different deployment settings. Existing benchmarks are reviewed along four axes: hallucination type, task, data construction and metrics. The authors find that standards are inconsistent and coverage is narrow. The work matters because it gives practitioners a structured map for choosing detectors. This is useful for any team, including those working with lower-resource languages, that builds or deploys LLM-based generation.

Key points

• Four-type taxonomy on two axes.
  • Input faithfulness covers semantic faithfulness hallucination (SFH) and contextual consistency hallucination (CCH).
  • Knowledge accuracy covers factual consistency hallucination (FCH), where the error comes from the model's internal memory, and external dependency hallucination (EDH), a new category for RAG settings where the retrieved knowledge is wrong or misused.
  • Each type gets a formal probabilistic definition and an example, e.g. naming Leonardo DiCaprio instead of Tom Hanks as the lead of Forrest Gump.
• Architectural causes. Three factors are discussed:
  • Model scale: small 7B/8B models such as LLaMA 7B tend to hallucinate more.
  • Unidirectional autoregressive modeling.
  • Attention dilution on long sequences.
• Pretraining causes. Data quality is examined through truthfulness, balance, bias, timeliness and domain specialization.
• Benchmarks cited. These include TruthfulQA, HaluEval, AutoHall, HALoGen, FactCHD, FaithBench, UHGEval, FELM, REALTIME QA and BAMBOO.
• Publication details. Journal of Computer Research and Development 63(1): 123–146, 2026; DOI 10.7544/issn1000-1239.202550069; funded by NSFC.