TY  - RPRT
TI  - Style-Debiased DPO: Updating LLM Knowledge with Factuality-Aware Synthetic Preference Data
AU  - Takayuki Yamamoto
AU  - Daisuke Kawahara
PY  - 2026
UR  - https://arxiv.org/abs/2609.16532
ID  - 2609.16532
ER  -