mlsanityData to 14 Oct 2026

Cross-Lingual Transfer of Evil Vectors Is Directionally Asymmetric

arXiv 2610.17880 ↗First posted 13 Oct 2026cs.CL cs.AIcited by 0

On the map: LLM safety and security → Safety alignment, refusal and guardrails

Abstract (arXiv)

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.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Safety alignment, refusal and guardrails, one of 12 directions in LLM safety and security. Builds on 20 of its 27 references that are in the map. First posted on arXiv 13 Oct 2026.

Builds on

20 of 27 references.

Closest ancestors · 1

Same direction, newest first

  1. 2024Refusal in Language Models Is Mediated by a Single Direction17 Jun

Other references · 19

  1. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,998
  2. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
  3. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
  4. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  5. 2023Steering Language Models With Activation Engineeringcited by 922
  6. 2025Persona Vectors: Monitoring and Controlling Character Traits in Language Modelscited by 345
  7. 2024Do Llamas Work in English? On the Latent Language of Multilingual Transformerscited by 281
  8. 2024Improving Instruction-Following in Language Models through Activation Steeringcited by 160
  9. 2021Systematic Inequalities in Language Technology Performance across the World's Languagescited by 120
  10. 2025Large Language Models Share Representations of Latent Grammatical Concepts Across Typologically Diverse Languagescited by 59
  11. 2024Probing the Emergence of Cross-lingual Alignment during LLM Trainingcited by 39
  12. 2024Multi-property Steering of Large Language Models with Dynamic Activation Compositioncited by 39
7 more
  1. 2025Refusal Direction is Universal Across Safety-Aligned Languagescited by 38
  2. 2025Middle-Layer Representation Alignment for Cross-Lingual Transfer in Fine-Tuned LLMscited by 29
  3. 2025Internal Value Alignment in Large Language Models through Controlled Value Vector Activationcited by 16
  4. 2025Can you map it to English? The Role of Cross-Lingual Alignment in Multilingual Performance of LLMscited by 12
  5. 2026PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebracited by 12
  6. 2025Aligning LLMs for Multilingual Consistency in Enterprise Applicationscited by 11
  7. 2026To Intervene or Not: Guiding Inference-time Alignment with Probabilistic Model Blendingcited by 2

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026What Models Express, Suppress, and Resist: Auditing Open-Weight LLMs with Persona Vectors
  2. 2024The Language Barrier: Dissecting Safety Challenges of LLMs in Multilingual Contexts
  3. 2026Hard Negatives Reveal What Easy Negatives Hide: Cross-Lingual Harmfulness Representations Degrade with Resource Tier Under Hard Negatives
  4. 2026Your Language Model Secretly Contains Personality Subnetworks
  5. 2026Lost in Translation? A Comparative Study on the Cross-Lingual Transfer of Composite Harms
  6. 2026The Illusion of Cross-Lingual Safety in Low-Resource Languages
  7. 2025The Geometry of Harmfulness in LLMs through Subconcept Probing
  8. 2026Do Linear Probes Generalize Better in Persona Coordinates?