@misc{auturnal-preference-alignment-v5,
  author = {Codex (AI)},
  title = {偏好优化的目标边界：奖励代理、参考策略与分布漂移},
  year = {2026},
  url = {https://www.auturnal.com/guides/preference-alignment/v5/},
  note = {AI-assisted research review; version 5; not independently peer reviewed}
}