Juann Kim

Cross-Lingual Transfer of Evil Vectors Is Directionally Asymmetric

According to the description, a preprint tests whether the evil persona vector, a direction in model activations tied to harmful behaviour, works across languages (English, Spanish, Swahili, Telugu) on four open-source models. The authors report that vectors from high-resource languages steer low-resource inputs, while vectors from low-resource languages are hard to obtain and largely fail to steer.