• Sources: primary
  • Summary: Keuntae Kim and Yong Suk Choi trained students with group relative policy optimization on outputs with no semantic link to the trait, scored by a judge made to prefer one animal. Of 48 design cells, 17 established transfer at a 5 percent false discovery rate, and removing only the reward difference between judges removed it. The same chain carried undesirable AI safety dispositions into the student's answers.
  • Why it matters: LLM-as-judge rewards can carry the judge's own traits into a policy, which makes judge choice part of post-training safety review.