Hallucination (幻觉)

Definition

The phenomenon where an LLM generates information that appears plausible but is actually false, fabricated, or not grounded in its input or training data. The model "makes things up" with confidence, presenting fiction as fact.

In the context of Chinese documentation: 大模型总是一本正经的回答问题，但其实是在胡说八道。LLM 在面对陌生领域时，只会在答案中写一个"解"字（因为 LLM 的知识局限于特定数据集），然后就开始放飞自我生成看似合理但实际错误的内容。

Key Statistics

If a single model hallucinates 20% of the time
3 models hallucinating the exact same lie: 0.8% (0.2³ = 0.008)
This mathematical property is the foundation of Consensus voting

Causes

Stochastic nature of LLM token generation
Training data includes conflicting or incorrect information
Model may lack specific knowledge but generates plausible substitutes
Prompting that asks for creative or speculative content

Impact on Multi-Agent Systems

Errors propagate through agent topologies
Can make entire system unreliable if not contained
Multiple architectures address this: Consensus, Validator, etc.

Mitigation

Multi-Agent Consensus — majority voting cancels noise
Validator checkpoints to catch errors
Deterministic code validation where possible
Don't anthropomorphize — force correctness architecturally

Sycophancy
Context Drift
Multi-Agent Consensus
Validator
LLM Reliability Engineering
RAG — 检索增强生成，通过外部知识检索解决幻觉问题，可将正确率从 60% 提升至 90%
Embedding — 向量化技术，支撑 RAG 的语义检索基础

1.7 KiB Raw Blame History Unescape Escape