Cross-Lingual Safety Transfer in Large Language Models

This direction is merged into another direction: cross-lingual safety disparity, cross-lingual safety transfer. Its last version stays readable and citable here.

Version 1 of 162 papers87 conceptsfrozen 2026-10-04

Large Language Models (LLMs) are typically "aligned" to be safe, helpful, and harmless through extensive Reinforcement Learning from Human Feedback (RLHF). However, this alignment is often heavily skewed toward English. When these same models are prompted in low-resource languages like Thai, Swahili, or Indonesian, they frequently exhibit a "safety gap," succumbing to harmful queries they would have easily refused in English.

Recent research has converged on a powerful observation: safety is not just a linguistic behavior, but a conceptual direction within a model's internal representation space. By leveraging this, researchers are developing ways to transfer safety guardrails across the linguistic divide without needing exhaustive human labels for every language on Earth.

The Geometry of Refusal

To understand why safety transfers, we must look at how LLMs represent abstract concepts. Researchers like Shaoyang Xu have shown that human values—such as fairness and morality—are often encoded as linear directions in a model's activation space. Because multilingual models map different languages to a shared semantic space, the "direction" for a harmful concept in English is often very close to the same concept in another language.

This shared geometry allows for Task Arithmetic. In a method called RESTA, Rishabh Bhardwaj and colleagues demonstrated that you can extract a "safety vector" by subtracting a base model's weights from a safety-aligned version. Adding this vector back into a model fine-tuned for a different language can restore safety guardrails almost instantly, suggesting that the structural logic of refusal is a portable trait.

Mechanisms of Cross-Lingual Alignment

If safety is a latent direction, how do we ensure the model follows it across all languages? Several strategies have emerged:

  • Latent Space Alignment: Instead of training on translated responses, Yuyan Bu and others proposed Multi-Lingual Consistency (MLC). This method forces the model's internal representations of harmful intent to be collinear across languages. If the model "sees" the same intent, it triggers the same refusal mechanism, regardless of the input language.
  • Internal Supervision: Rather than relying on noisy machine-translated data, the MPO framework by Weixiang Zhao uses the model's own robust English safety signals to calibrate its behavior in other languages. The model essentially acts as its own teacher, using its high-resource knowledge to anchor its low-resource behavior.
  • Decoupled Refusal: A common failure mode is "refusal position bias," where models only know how to refuse at the start of a sentence. Youliang Yuan introduced DeRTa, which trains models to pivot to a refusal at any point in a sequence. This capability, once learned, generalizes across languages, hardening the model against sophisticated "jailbreak" attempts that try to hide harmful intent mid-sentence.

Bridging the Linguistic Divide

One of the most effective ways to facilitate this transfer is to explicitly link languages during training. Haneul Yoo developed Code-Switching Curriculum Learning (CSCL), which progressively mixes languages (e.g., English and Korean) during fine-tuning. This creates an explicit mapping between the two, allowing safety constraints to flow from the well-aligned source to the target.

Similarly, the SELF-DEFENSE framework by Yue Deng uses the LLM itself to generate synthetic multilingual safety data. By repurposing the model's internal knowledge, developers can automate the alignment process for dozens of languages simultaneously.

The Trade-offs: Fluency and Culture

While cross-lingual transfer is powerful, it is not a silver bullet. Arianna Bisazza has highlighted a significant "language confusion" trade-off: aggressive safety tuning in English can cause a model to lose fluency in other languages or default to English when issuing refusals.

Furthermore, safety is often culturally dependent. A model aligned on Western values might not understand the specific safety risks in an Indonesian context. To address this, Muhammad Falensi Azmi and Raviraj Joshi have pushed for "culturally grounded" datasets like IndoSafety and CultureGuard. These methods adapt safety prompts to regional contexts—such as local scams or social norms—ensuring that the transferred safety is not just linguistically accurate, but socially relevant.

The emerging consensus is that while the mechanism of refusal is universal and transferable, the context of what constitutes harm must still be grounded in the specific culture of the language being spoken.

Papers in this version

Showing 60 of 62 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_cross_lingual_safety_transfer_in_large_l_4a170502,
  title = {Cross-Lingual Safety Transfer in Large Language Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-lingual-safety-transfer-in-large-language-models/rdr_4a170502c07e435eb99dda969aa187a4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_cross_lingual_safety_transfer_in_large_l_4a170502_v1,
  title = {Cross-Lingual Safety Transfer in Large Language Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-lingual-safety-transfer-in-large-language-models/rdr_4a170502c07e435eb99dda969aa187a4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}