aiqbench analysis
This research investigates the cross-lingual transferability of "evil persona vectors" within LLM residual streams, revealing a directional asymmetry where high-resource language vectors effectively steer low-resource models while the reverse fails. The findings indicate that a shared underlying direction for harmful behaviors exists across languages, but its reliable extraction is contingent on training data richness rather than linguistic capability alone. These insights are critical for understanding jailbreak vulnerabilities and developing robust unlearning strategies for multilingual foundation models.
Cross-Lingual Transfer of Evil Vectors Is Directionally Asymmetric
arXiv:2610.17880v1 Announce Type: cross Abstract: Persona vectors are linear directions in the residual-stream activations of LLMs that encode behavioral traits and enable fine-grained control over model behavior. Among them, the evil persona vector is particularly relevant to understanding harmful behavior, jailbreak vulnerabilities, and the unlearning of harmful knowledge. Yet whether this vector reflects a genuinely cross-lingual concept, rather than an artifact of its extraction language, remains untested. We conduct a systematic evil-vector transfer study across two high-resource languages (English, Spanish) and two low-resource languages (Swahili, Telugu) on four open-source instruction-tuned LLMs, and find that transfer is directionally asymmetric. Evil vectors are harder to obtain reliably from low-resource languages, and the resulting vectors are largely ineffective at steering, even within their own language. In contrast, vectors extracted from high-resource languages transfer effectively to low-resource inputs while largely preserving general knowledge, as measured by MMLU. Geometric analysis further supports this pattern: evil vectors from the two high-resource languages are closely aligned, whereas low-resource vectors align only weakly with this shared direction. These results suggest that a shared cross-lingual evil direction exists and, once reliably extracted from a high-resource language, can be reused to steer behavior in low-resource languages without additional per-language data collection.
Abstract:Persona vectors are linear directions in the residual-stream activations of LLMs that encode behavioral traits and enable fine-grained control over model behavior. Among them, the evil persona vector is particularly relevant to understanding harmful behavior, jailbreak vulnerabilities, and the unlearning of harmful knowledge. Yet whether this vector reflects a genuinely cross-lingual concept, rather than an artifact of its extraction language, remains untested. We conduct a systematic evil-vector transfer study across two high-resource languages (English, Spanish) and two low-resource languages (Swahili, Telugu) on four open-source instruction-tuned LLMs, and find that transfer is directionally asymmetric. Evil vectors are harder to obtain reliably from low-resource languages, and the resulting vectors are largely ineffective at steering, even within their own language. In contrast, vectors extracted from high-resource languages transfer effectively to low-resource inputs while largely preserving general knowledge, as measured by MMLU. Geometric analysis further supports this pattern: evil vectors from the two high-resource languages are closely aligned, whereas low-resource vectors align only weakly with this shared direction. These results suggest that a shared cross-lingual evil direction exists and, once reliably extracted from a high-resource language, can be reused to steer behavior in low-resource languages without additional per-language data collection.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2610.17880 [cs.CL]
(or arXiv:2610.17880v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2610.17880 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 294 words
Related coverage
Cross-Lingual Steering for Figurative Language Generation
arXiv · Oct 14, 2026
Language Steering for Multilingual In-Context Learning
arXiv · Oct 15, 2026
Cross-lingual Functional Vectors for Emotion Detection in Large Language Models
arXiv · Sep 1, 2026
Asymmetric Hierarchical Anchoring for Robust Audio-Visual Cross-Modal Generalization
arXiv · Jul 28, 2026