论文进阶
hallucination
幻觉。指大模型生成的内容读起来流畅合理，但实际上没有事实依据、检索证据或输入支撑。在医疗、法律等场景里尤其危险。
English Definition
"In LLM/AI contexts, an output that is fluent and plausible but is not supported by the input, retrieved evidence, or factual reality."
用法说明 · 针对中文母语者
在设计/AI 行业里，hallucination 已成标准术语；中文文献也常直接借用。论文中常与 RAG、grounding、faithfulness 配套讨论。
常见错误
- ·把 hallucination 等同于"胡说八道"——它通常听起来非常自信、专业
- ·以为加 RAG 就完全没幻觉
真实用例 · 3 条
"general-purpose models are prone to providing unsupported facts and often lack updated domain-specific knowledge"
论文Pipeline extracted — review needed·通用 LLM 在医疗场景的问题
"while RAG significantly reduces hallucinations, it does not eliminate them entirely"
论文Pipeline extracted — review needed·RAG 的边界
"the deployment of general-purpose LLMs in healthcare is hindered by their propensity to 'hallucinate' facts, where evidential support is lacking"
论文Pipeline extracted — review needed·Related Works 中复述 LLM 缺陷
由 pipeline 自动采集，待人工 review