Extracting Persona Subspaces Through Iterative Nullspace Projection For Modulation
The paper argues that persona control works better when a model’s persona is treated as a subspace, not a single steering direction.
Malik and ElSherief introduce PaSS, an inference-time method that uses Iterative Nullspace Projections to peel out multiple persona-specific directions from a model’s latent space. The setup is modulation: amplifying or suppressing a trait already present in the content, rather than adding a persona from scratch. They evaluate six personas across MATH-500, TinyAlpaca, GSM8K, and IFEval, reporting stronger modulation than single-direction additive methods while preserving content fidelity. The authors also inspect individual directions in each subspace to identify distinct aspects of persona behavior. ArXiv · AI/CL/LG's note
Malik and ElSherief introduce PaSS, an inference-time method that uses Iterative Nullspace Projections to peel out multiple persona-specific directions from a model’s latent space. The setup is modulation: amplifying or suppressing a trait already present in the content, rather than adding a persona from scratch. They evaluate six personas across MATH-500, TinyAlpaca, GSM8K, and IFEval, reporting stronger modulation than single-direction additive methods while preserving content fidelity. The authors also inspect individual directions in each subspace to identify distinct aspects of persona behavior. ArXiv · AI/CL/LG's note
score 4