模型对你专业水平的内部估计，会引导它如何回答你，以及是否"录用"你。 大语言模型携带一个关于用户能力的单一方向，同时驱动回答的复杂度 这篇论文表明该估计以因果方式中介了行为，但并未表明依附其上的 demographic 差异会可靠地在输出中表现为歧视。 – arxiv. org/abs/2608.20347 标题："Who Do Language Models Think Is Competent? A Mechanistic Analysis of Occupational Bias"
A model's internal estimate of your expertise steers how it answers you and whether it hires you.
Language models carry a single direction for user competence that drives both answer complexity
This paper shows that estimate causally mediates behavior, but not that the demographic differences riding on it reliably surface as discrimination in the output.
– arxiv. org/abs/2608.20347
Title: "Who Do Language Models Think Is Competent? A Mechanistic Analysis of Occupational Bias"
来源：@rohanpaul_ai · x.com