Embracing Anisotropy: Turning Massive Activations into Interpretable Control Knobs for Large Language Models.

Youngji Roh, Hyunjin Cho, Jaehyung Kim

Browse the full ACL paper archive.