幻觉的成因Why Hallucination Happens
Why Hallucination Happens
为什么模型会一本正经地编：不是它想骗你，是训练目标使然。Why models fabricate with a straight face: not deception, but the training objective.
语言模型本质是逐词概率预测器，训练要求它永远给出下一个词的猜测。
Anthropic 的研究页直言：训练鼓励幻觉。模型于是会编出貌似合理但不真的回答。
学界常把幻觉定义为「生成内容不被输入来源支持或与事实不符」。
流畅自然的表象与不忠实并存，这正是它难被发现的原因。
A language model is a word-by-word probability predictor; training requires it to always guess the next word.
Anthropic's research page states plainly that training incentivizes hallucination, so the model confabulates plausible-but-untrue responses.
The standard definition: generated content not supported by the input source or contrary to fact.
Fluent, natural surface coexists with unfaithfulness — that is exactly why it is hard to spot.
知道成因，才知道哪些缓解手段有效、哪些只是安慰。给出处、允许说不知道属于前者。Knowing the cause tells you which mitigations actually work (provide sources, allow 'I don't know') and which are just reassurance.
回到名词地图看第4站的其他 35 个词 →Back to the map: the other 35 terms in station 4 →