拒答Refusal
Refusal
模型对某些请求直接拒绝回答：安全对齐的直接产物。The model flatly declines certain requests: a direct product of safety alignment.
对齐训练的目标就是让模型拒绝产出不当内容。
Zou 等人的攻击研究以此为前提：他们的攻击目标是让模型「给出肯定回答而非拒绝」。
拒答是护栏起作用的信号，也是误伤正常请求时的摩擦来源。
同一个请求在不同家模型上拒与不拒差异很大，这是产品间的真实差异。
Alignment training aims precisely at making models refuse to produce disallowed content.
Zou et al.'s attack research takes this as its premise: their attack aims for 'an affirmative response rather than refusing to answer'.
Refusal is a signal the guardrails are working — and a source of friction when legitimate requests get caught.
The same request refused by one vendor and answered by another is a real product difference.
遇到拒答，你该知道这不是「模型坏了」，而是对齐在工作。评估一家模型时，拒答边界的松紧是选型维度。When refused, you should know the model isn't 'broken' — alignment is working. When evaluating models, refusal boundaries are a selection dimension.
回到名词地图看第2站的其他 35 个词 →Back to the map: the other 35 terms in station 2 →
- Zou 等《Universal and Transferable Adversarial Attacks on Aligned Language Models》（对齐与拒答的表述出处）Zou et al., Universal and Transferable Adversarial Attacks (alignment/refusal context) ↗