Tag
This paper introduces an unsupervised, training-free approach to discover prompt-conditional stylistic axes in LLM hidden activations using sampling and PCA, validated through human studies.