从最优传输角度训练奖励模型：让 RLHF 学会「忽略错误偏好」丨ICML 2026

<section data-mpa-template="t" mpa-data-temp-power-by="yiban.io" mpa-data-temp-type="image" mpa-from-tpl="t" data-mpa-action-id="mmix4tgk1t7p" data-pm-slice="5 10 []" sty