作者位次 #5/12ACL 2026Poster· Natural Language Processing
DARM : Distribution-Aware Reward Modeling by Alleviating Biases from Low Preference-Context Dependency Data
Shaofan Liu · Guoqiang Zhang · Shihan Dou · Huiyuan Zheng · Yiming Zhou · Junjie Ye · Shaowen Wang · Shichun Liu · Jiazheng Zhang · Tao Gui · Qi Zhang · Xuanjing Huang
摘要
Reward models (RMs) are the surrogate objectives in reinforcement learning from human feedback (RLHF), and their scores directly steer policy optimization. We show that standard RM training is vulnerable in data subsets where response quality depends only weakly on the context: such instances enc...