hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
grounding — Tying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
Why it matters
It gives a precise taxonomy of hallucinationWhen a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.Full definition → (in-context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → vs. extrinsic) and surveys the actual detection and mitigation methods, so you can design evals and factuality guardrailsThe checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.Full definition → around what your model genuinely doesn't know versus what it fabricates.
Key quotes
“To avoid hallucination, LLMs need to be (1) factual and (2) acknowledge not knowing the answer when applicable.”
“Equally importantly, when the model does not know about a fact, it should say so.”
articleOpen challenges in LLM researchChip Huyen
articleUnified Hallucination Fuzzing for Multimodal Large Language ModelsPengfei Zhou, Jiajun Song, Zhiwei Tang, Yixing Ma, Xiaopeng Peng, Donghui Si, Yuhang Xu, Huiqi Song, Yiyuan Miao, Yichen Qian, Weihua Chen, Wangbo Zhao, Bohan Zhuang, Jiasheng Tang, Yang You