A new preprint on arXiv examines how persona vectors—linear directions in a language model's residual-stream activations that encode behavioral traits—behave across languages. These vectors allow fine-grained control over model behavior, and the 'evil' persona vector is especially relevant to safety. The paper's central finding is that cross-lingual transfer of this vector is directionally asymmetric: transferring from one language to another does not behave the same as the reverse transfer.

The authors do not claim that the vector fails to transfer, but that the strength or effect depends on the source and target languages. This asymmetry matters because safety interventions that rely on steering vectors in one language may not work as intended in another. The paper is a single preprint, so the results have not yet been peer-reviewed, but they point to a concrete complication in building multilingual models with consistent behavior controls.