Hallucination
Core Facts
In multi-Agent scenarios, hallucination problems are cascadingly amplified: the erroneous output of one Agent may become the input of the next, causing errors to accumulate continuously throughout the chain.
Hallucination refers to content generated by a model that appears plausible but is actually incorrect or fabricated, and is the most intractable technical problem with LLMs.
Strategies for mitigating LLM hallucination include retrieval-augmented generation (RAG), chain-of-thought prompting, and requiring the model to output confidence scores.
Timeline (last 90 days)
In a RAG pipeline, context carries factual information, so a low hallucination rate is an advantage for a model; in open-domain question answering scenarios without retrieval support, a low hallucination rate becomes a shortcoming (higher refusal rate, lower coverage).
Hallucination is an inherent property of the probabilistic generation mechanism rather than a defect of individual models, and cannot be mitigated by increasing model parameters or prompt words.
The supervisor and the supervised share the same class of failure modes: if the Agent makes a wrong decision due to hallucination, the LLM Supervisor may likewise let that decision through due to hallucination.
If an agent is given full permissions to directly send emails or modify data, a single hallucination will cause immediate harm, such as sending an incorrect email to a customer.
Having a second model perform a structured review against a rules checklist in the role of a critic can significantly reduce the probability of both models simultaneously producing the same hallucination.
Hallucinations occur at a significantly higher rate in scenarios such as querying niche or long-tail knowledge, requiring precise numbers and dates, involving recent events beyond the training cutoff date, and generating academic content with references.
Grounding prompts that instruct the model to answer only based on provided documents and cite the relevant passages can greatly reduce the probability of fabrication, but cannot eliminate it entirely.
Hallucination refers to the phenomenon where a large language model generates content that appears plausible but is actually incorrect or fabricated.
Mainstream approaches in the industry for addressing AI hallucinations include mandating citation of original data sources, introducing RAG (Retrieval-Augmented Generation) mechanisms, and adding confidence annotations at the output layer.
Agents may execute destructive operations far beyond what was intended due to hallucination or prompt injection attacks.
7 more timeline events
All Facts (20)
In multi-Agent scenarios, hallucination problems are cascadingly amplified: the erroneous output of one Agent may become the input of the next, causing errors to accumulate continuously throughout the chain.
90%VerifiedHallucination refers to content generated by a model that appears plausible but is actually incorrect or fabricated, and is the most intractable technical problem with LLMs.
80%VerifiedStrategies for mitigating LLM hallucination include retrieval-augmented generation (RAG), chain-of-thought prompting, and requiring the model to output confidence scores.
80%VerifiedThe root cause of LLM hallucination lies in the autoregressive generation mechanism—the model predicts the next Token based on a probability distribution rather than genuine logical reasoning.
80%VerifiedLLM幻觉（Hallucination）是大语言模型的已知问题，会导致AI生成不准确的内容
80%VerifiedLLMs have a hallucination problem, in which they generate factually incorrect information with high confidence.
80%Verified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%VerifiedLarge language models carry a hallucination risk in numerical reasoning and may produce errors in scenarios such as aggregate calculations, time-range filtering, or multi-table joins.
75%VerifiedRAG technology can break through the temporal limitations of model training data and significantly reduce the probability of hallucination occurrences.
75%VerifiedLLM产生幻觉的深层机制在于模型被优化为生成'听起来合理'的文本而非'事实准确'的文本，模型没有独立的知识检索模块
70%VerifiedLLMs have a hallucination characteristic, potentially generating code that appears plausible but actually contains logical errors or functional defects with confidence, without proactively flagging it.
70%VerifiedAI幻觉是大语言模型生成机制的内在产物，模型目标是生成听起来合理的内容而非绝对真实的内容
65%VerifiedThe root cause of hallucination lies in the fact that the model's training objective optimizes for generating fluent and coherent text rather than generating truthful text.
65%VerifiedLarge models are prone to 'hallucination' when handling professional domain questions, meaning they confidently provide incorrect answers.
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%UnverifiedHallucination refers to LLMs generating plausible-sounding but factually incorrect information
90%Unverified幻觉（Hallucination）分为内在幻觉和外在幻觉两类，检测技术路线包括基于知识库的事实核查、基于模型不确定性的检测和基于多次采样一致性的方法
90%UnverifiedReducing hallucination rates typically requires optimization across training methods such as RLHF and Retrieval-Augmented Generation (RAG) architectures
85%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大语言模型幻觉（Hallucination）在多轮对话或多步推理的Agent场景中危害被成倍放大，形成「幻觉级联」（Hallucination Cascade）
85%