LLM Hallucination
The phenomenon in which large language models generate content that appears plausible and fluent but is factually incorrect, unsupported, or inconsistent with the input. Causes include noise in training data, probabilistic generation mechanisms, and a lack of self-awareness of knowledge boundaries; it can be mitigated through methods such as RAG and knowledge-base anchoring.
Core Facts
大模型幻觉的技术本质在于模型在训练时学习的是语言模式的概率分布而非事实本身
In multi-step automation scenarios, single-step errors caused by model hallucinations are amplified exponentially along the task chain.
In Agent chain-call scenarios, a hallucination at one step is inherited as fact by subsequent steps, causing errors to amplify and propagate along the pipeline.
Timeline (last 90 days)
As of now, large models still cannot truly understand every human behavior, and hallucinations and instability still exist.
Other lawyers on the team reported that five models gave five different legal opinions on the same legal question, showing that hallucinations are common.
Because large models hallucinate and produce unstable outputs, agents in real-world engineering rely more on rules for execution, with most rules configured through prompts.
In Agent chain-call scenarios, a hallucination at one step is inherited as fact by subsequent steps, causing errors to amplify and propagate along the pipeline.
幻觉（Hallucination）指大语言模型生成看似合理但实际错误或无依据的内容，是当前所有LLM的共性问题，根源在于模型本质上是基于概率的文本生成器而非知识检索系统
All Facts (10)
大模型幻觉的技术本质在于模型在训练时学习的是语言模式的概率分布而非事实本身
80%VerifiedIn multi-step automation scenarios, single-step errors caused by model hallucinations are amplified exponentially along the task chain.
75%VerifiedIn Agent chain-call scenarios, a hallucination at one step is inherited as fact by subsequent steps, causing errors to amplify and propagate along the pipeline.
65%Verified大模型幻觉是指语言模型生成看似合理但实际不正确或不存在的内容
65%Unverified幻觉（Hallucination）指大语言模型生成看似合理但实际错误或无依据的内容，是当前所有LLM的共性问题，根源在于模型本质上是基于概率的文本生成器而非知识检索系统
85%Unverified添加行号后模型能更准确地定位代码位置，减少行号幻觉
75%UnverifiedAs of now, large models still cannot truly understand every human behavior, and hallucinations and instability still exist.
50%UnverifiedOther lawyers on the team reported that five models gave five different legal opinions on the same legal question, showing that hallucinations are common.
50%UnverifiedBecause large models hallucinate and produce unstable outputs, agents in real-world engineering rely more on rules for execution, with most rules configured through prompts.
50%Unverified级联失败是指单次工具调用错误会污染后续上下文，导致模型在错误的中间状态上继续推理，最终使整个任务偏轨
50%