Attention Limited Reward Learning

Wenqian Xing

Pairwise human comparisons are a primary interface through which modern AI systems learn human preferences. RLHF and related alignment pipelines typically model such comparisons wi…