TY  - RPRT
TI  - Compositional Bias Control in Large Language Models: Preference Learning Fails, Supervision Succeeds
AU  - Atij Mahesh
PY  - 2025
UR  - https://arxiv.org/abs/2510.22084
ID  - 2510.22084
ER  -