偏好对齐RLHF / Alignment
RLHF / Alignment
解决什么问题What problem it solves
让模型的输出符合人类偏好，而不是只会续写文本。Makes a model's output match human preference, rather than merely continuing text.
这是什么What it is
用人类对多个候选回答的排序来训练模型，让它更倾向于给出有帮助、无害、诚实的回答。
Training a model on human rankings of several candidate answers so it tends to be helpful, harmless and honest.
为什么需要知道Why it matters
它解释了为什么模型有时候会「不肯回答」或「过度客气」。那是训练出来的，不是技术限制。It explains why a model sometimes refuses or is excessively agreeable — that is trained behaviour, not a technical limit.