Huang et al. — A Survey on Hallucination in Large Language Models (2023)

2 min read

A Survey on Hallucination in Large Language Models

Authors: Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, Ting Liu (Harbin Institute of Technology) Publisher: ACM TOIS Canonical URL: https://arxiv.org/abs/2311.05232

Summary

The current canonical survey of hallucination in LLMs. Introduces a taxonomy that distinguishes LLM-era hallucination from prior task-specific NLG hallucination, surveys causes, detection methods, and mitigation strategies, including the limitations of retrieval-augmented generation (RAG) as a fix.

Key Contributions

  1. Updated taxonomy. Splits hallucination into factuality (the claim is wrong about the world) and faithfulness (the claim is unsupported by provided source material) — important distinction often conflated.
  2. Causes are multi-layered. Pre-training data quality, training objective, RLHF artifacts, decoding randomness, and prompting all contribute.
  3. Detection methods. Token-level uncertainty, consistency checks, retrieval-based verification, model-as-judge — each with trade-offs.
  4. RAG is not a cure. Even with retrieval, LLMs hallucinate when the retrieved context is irrelevant, contradictory, or absent — and they often fail to abstain when retrieval fails.
  5. Knowledge boundaries. Models do not reliably know what they do not know (compare with Kadavath et al. 2022).

Relevance to This Wiki

  • Primary reference for Hallucination concept page.
  • Faithfulness/factuality split is important for H7 design: epistemic labeling can plausibly improve faithfulness (the model knows what it was given) but cannot fully solve factuality (the model doesn’t always know what’s true).
  • RAG-is-not-a-cure finding is critical for any agentic system relying on retrieval as a hallucination guard — SATs may still be needed even when sources are provided.

See Also

Graph View

Backlinks