Large Language Models in Resolving Contextual Knowledge Conflicts
A simple training-free, label-free steering method that steers activations to encourage a more comprehensive incorporation of evidences for better conflict resolution is proposed and consistently improves accuracy on reasoning tasks and generates higher-quality, more balanced summaries for summarization tasks.
Xin-Ye Yang, Zhen-Yang Liu, Rui-Si Li et al.
· 0 citations