Across 29 language models and 177 occupations (covering nearly half of U.S. employment), post-training alignment amplifies demographic effects in hiring: female and Black candidates gain advantages while disabled candidates face larger penalties; these effects rival several months to a year of education. Pre-trained models show smaller effects; alignment substantially increases sensitivity to credentials and shifts demographics in model representations, with disability underrepresented in training data and negatively encoded in deeper layers .
一、元数据与概览(引言与研究概览)
图像提示(溯源用的图像引用,便于后续逐图解读)
二、逐节深度解读(逐章精读与剖析)
三、图表深度解读(逐图分析与溯源)
四、估值分析(定量效果与经济意义的解读)
五、风险因素评估(局限性与潜在偏差)
六、批判性视角与细微差别(审慎评估)
七、结论性综合(综合要点与作者立场)
附:方法要点与数据可获得性(简要回顾,便于理解分析的基础)
结论性综合(最终洞见)
补充说明(溯源清单,便于后续追溯)
若需要,我可以将以上分析扩展成逐表逐图的对照表,逐项标注每个数据点的具体数值与对应页码,便于快速溯源与后续撰写相关稿件。
2605.13866.pdf