Unverified50% confidenceFactExact time
大模型幻觉可分为事实性幻觉（捏造不存在的事实）和忠实性幻觉（偏离用户指令或上下文语义）两类，前者与训练数据覆盖有关，后者与RLHF对齐训练质量相关
1
Sources
50%
Confidence
Long-term
Relevance
7/14/2026
First Seen
Sources
零基础学AI大模型：避开三大误区的完整学习路径
Related Claims
UnverifiedLoRA训练中，简单提示词测出的高还原度可能是过拟合导致的假象，需引入复杂新颖场景验证真实泛化能力71% similarUnverifiedModel-based reinforcement learning can leverage physical priors to conduct imagination rollouts, reducing the number of interactions with the real environment and avoiding dangerous states.63% similarUnverified大模型在训练时学习的是静态知识模式，难以感知真实业务场景中界面元素属性值动态变化的运行时不确定性63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/507659API
curl https://kongchang.com/api/v1/knowledge/claims/507659MCP
get_claim(id=507659)