输出越界与拒答Output Overreach and Refusal
Output Overreach and Refusal
看清模型什么时候会拒绝、什么时候会多做，知道自己该期待哪一种反应。See which requests a model will decline, which it will answer only partly, and when it does more than you asked for.
拒答是模型对某类请求给出不合作的回复。越界是相反方向：它在你没要求的范围里多做了事，比如顺手改了你没打算动的文件。
OpenAI 的 Model Spec 把模型要服从的指令排成平台、开发者、用户三级，里面既规定了拒绝时保持中立和简短，也写了在可以部分满足时给可用的那部分，而不是一拒了之。
拒答这件事本身是可被调制的。Arditi 等人 2024 年的研究报告：去掉某个内部方向后模型不再拒绝有害指令，加上这个方向后连无害提问也会被拒绝。
同一个请求在不同产品上拒与不拒不一样。这是训练目标和策略配置的差别，不是你的提示词写错了。
Refusal is the model declining to cooperate with a request. Overreach runs the other way: it does work you did not ask for, such as editing a file you had not meant to touch.
OpenAI's Model Spec sorts the instructions a model must obey into platform, developer and user tiers, tells it to refuse neutrally and briefly, and allows a partial answer when the full request cannot be met.
Refusal itself turns out to be a tunable behaviour. Arditi and colleagues reported in 2024 that removing one internal direction stopped models from refusing harmful requests, while adding that same direction produced refusals even on harmless ones.
The same request may be refused by one product and answered by another. That is a difference in training objective and policy configuration, not a sign that your prompt was badly written.
遇到拒绝，你要知道下一步是换个说法拿部分回答，还是换工具。遇到越界，你要知道那是越界，不是贴心。把两个方向分开，能省掉大量返工和误判。When you get a refusal you need to know the next move: rephrase to recover a partial answer, or switch tools. When the model has done extra work you need to recognise it as overreach rather than helpfulness. Telling the two apart saves a lot of rework.
回到名词地图看第2站的其他 41 个词 →Back to the map: the other 41 terms in station 2 →