arxivcs.LG2026-07-17
TD-DPO: Difference-Aware Preference Optimization for Mitigating Sycophancy in Clinical Autism Intervention Dialogue
Shuzhong Lai, Junhong Lai, Chenxi Li, Qing Zhou, Haifeng Li, Gang Pan, et al.
The sycophancy of large language models can increase the safety risk in intervention dialogue for autistic children. Supervised fine-tuning can somewhat reduce sycophancy, but relying solely on positive examples is often insufficient to identify and correct failure patterns. We o…