Measuring, Localizing, and Ablating Alignment Signatures in LLMs

Aniket Anand, Janvijay Singh, Zhewei Sun +2

Aligned language models often exhibit a recognizable AI-like style, yet its connection to post-training and internal representations remains poorly understood. In this work, we stu…