When we align a Large Language Model (LLM) to be safe, we typically do so using high-quality instruction and preference data in a single, dominant language like English. Yet, remarkably, a model safety-tuned exclusively in English will often refuse harmful prompts written in French, Hindi, or Indonesian. This phenomenon—cross-lingual safety transfer—suggests that safety alignment is not merely a set of surface-level text associations. Instead, the model learns ethical constraints at a conceptual level that transcends individual languages.
The Universal Refusal Vector
To understand why safety generalizes across languages, we must look at how LLMs represent the act of refusal internally. Wang et al. discovered that the internal logic governing whether a model complies with or refuses a request is mediated by a single, language-agnostic direction in its activation space.
By isolating this "refusal vector" using English prompts, the authors could manipulate or bypass safety filters across 13 entirely different languages and writing systems. This demonstrates that regardless of the input language, the model routes harmful intents through a centralized, universal geometric bottleneck to trigger a refusal.
[Harmful Prompt in English] ----\
\---> [Shared Intent Space] ---> [Universal Refusal Vector] ---> "I cannot fulfill this request."
[Harmful Prompt in Swahili] ----/
This high-level capability is also evident in how models manage conversational transitions. Yuan et al. showed that training a model to recognize risks and transition into a refusal state at any point in a sequence generalizes across multiple languages without requiring language-specific safety data. The ability to recognize and interrupt harmful generation is a structural feature of the model's reasoning, not a linguistic one.
Shared Semantic Spaces and Intent Alignment
Why do diverse languages converge on the same internal concepts? During multilingual pre-training, LLMs map different languages into a shared latent space where semantically equivalent concepts align.
Xiao et al. demonstrated this geometric consistency by showing that a model's "knowledge boundary"—its awareness of what it does and does not know—is encoded linearly across languages. By applying simple, training-free linear transformations to align these representation spaces, safety and truthfulness insights can be transferred from high-resource to low-resource languages.
Similarly, the authors of Align Once, Benefit Multilingually leveraged this property by introducing a latent space alignment loss that forces the internal representations of harmful intent across different languages to be collinear. Because the model learns to recognize the underlying intent in a shared semantic space rather than focusing on language-specific tokens, safety behaviors naturally extend to entirely unseen languages zero-shot, without requiring translated safety responses during training.
Global Suppression via Preference Tuning
This conceptual generalization is also evident in how parameter updates affect the model's internal layers. When performing Direct Preference Optimization (DPO) to mitigate toxicity, training on English data alone can suppress toxic activations globally. Li et al. found that the model's internal MLP layers represent toxic concepts in a language-agnostic manner. Consequently, preference tuning in a single source language modifies these shared conceptual pathways, yielding zero-shot detoxification across dozens of target languages.
Furthermore, these safety characteristics are remarkably robust and consistent within the model's weights. Bhardwaj et al. showed that safety alignment forms consistent parameter directions. By using task arithmetic to add a "safety vector" back into the weights of a compromised model, they could restore safety across multiple languages simultaneously, confirming that the underlying mechanisms of safety are decoupled from the linguistic features of the fine-tuning data.
Asymmetries and the Resource Gap
While cross-lingual transfer is powerful, it is not perfectly symmetrical. The success of safety propagation depends heavily on how well a language is represented in the original pre-training corpus. Bisazza et al. observed that high-resource languages inherit safety properties from English-centric alignment far more effectively than low-resource languages, though this can sometimes come at the cost of target-language fluency.
Additionally, Xu et al. identified a "unidirectional transfer" dynamic: internal concept vectors extracted from dominant languages like English can easily steer or manipulate behaviors in low-resource languages, but the reverse is rarely true. This asymmetry leaves low-resource languages vulnerable to adversarial exploits and multilingual jailbreaks, as highlighted by Deng et al..
Bridging the Cross-Lingual Safety Gap
To overcome these resource disparities and secure low-resource languages, researchers have developed targeted alignment strategies. One approach is to explicitly tighten the representational alignment between languages during training. Yoo et al. demonstrated that using a code-switching curriculum—mixing source and target languages within the training data—smooths the transfer of safety guardrails and reduces vulnerability to multilingual jailbreaks.
Alternatively, developers can exploit the model's own latent capabilities. Deng et al. showed that a model's internal multilingual knowledge can be used to synthetically generate its own safety data, automating the alignment process across underrepresented languages.
When combined with multi-phase instruction tuning, as shown by Cahyawijaya et al. in the development of Indonesian models, or preference optimization frameworks like those explored by Aakanksha, models can maintain robust safety boundaries globally without requiring exhaustive, language-specific human feedback.
Key Foundational Work
Aurora-M: Open Source Continual Pre-training for Multilingual Language and Code (2025) — contributed 2 concept elements to this direction.
Cendol: Open Instruction-tuned Generative Large Language Models for Indonesian Languages (2024) — contributed 2 concept elements to this direction.
DIPLomA: Efficient Adaptation of Instructed LLMs to Low-Resource Languages via Post-Training Delta Merging (2025) — contributed 2 concept elements to this direction.