Tag
This paper provides a mathematical analysis of the Jacobian lens (J-lens) for reading verbalizable representations from language models, identifies its sparse causal geometry, and proposes improvements to enhance concept readout.