TY  - RPRT
TI  - Decomposing the Delta: What Do Models Actually Learn from Preference Pairs?
AU  - Chia-Hsuan Lee
AU  - Mingyang Zhou
AU  - Renkun Ni
AU  - Zelei Cheng
AU  - Sihui Dai
AU  - Supriyo Chakraborty
AU  - Shixiong Zhang
AU  - Sambit Sahu
AU  - William Campbell
PY  - 2026
UR  - https://arxiv.org/abs/2604.08723
ID  - 2604.08723
ER  -