Xinying Qian

Rewarding Better Thinking for LLM Preference Alignment

Xubo Liu, Wenya Guo, Ruxue Yan, Xinying Qian, Ying Zhang

LLM preference alignment aims to optimize models toward human preferences across diverse user instructions. Reinforcement learning has become a major post-training approach for this goal, but existing proxy rewards are often outcome-level, mainly evaluating the final response whi…Read more