Technical Summary: LLM Safety Alignment in Low-Resource Languages

Problem Statement

While Large Language Models (LLMs) have achieved significant progress in safety alignment for high-resource languages (particularly English), their safety guarantees degrade substantially in low-resource and multilingual settings. Current alignment techniques, such as Reinforcement Learning from Human Feedback (RLHF) and safety fine-tuning, rely heavily on large-scale annotated datasets and human feedback pipelines that are scarce or non-existent for underrepresented languages. Consequently, safety behaviors learned in English do not transfer reliably; models often comply with harmful instructions when prompted in low-resource languages (LRLs) or fail to recognize culturally specific harms. This creates a "multilingual safety gap" where speakers of underrepresented languages face weaker protections against toxic, biased, or dangerous content.

Methodology

This paper presents a Systematic Literature Review (SLR) conducted according to the PRISMA 2020 guidelines to synthesize existing research on LLM safety alignment in low-resource languages.

• Data Collection: The authors queried Semantic Scholar, arXiv, and OpenAlex using three keyword sets: technology (e.g., LLM, instruction tuning), safety (e.g., alignment, jailbreak, RLHF), and linguistic scope (e.g., low-resource, African languages, specific languages like Hausa and Swahili).
• Selection Process: From an initial pool of ~1,500 papers, duplicates were removed, leaving 1,300. A Python-based filter and an LLM-based screening process (using Claude Sonnet) narrowed this to 70 candidates. After manual verification, 50 relevant studies were selected for the final analysis.
• Scope: The review focuses on safety alignment in multilingual and low-resource contexts, explicitly excluding purely English and non-empirical works.
• Research Questions (RQs): The review is structured around four questions:
  1. How are safety methods adapted for LRLs?
  2. What safety risks and cultural harms emerge in multilingual settings?
  3. What datasets and benchmarks exist for LRL safety?
  4. What factors affect the cross-lingual transfer of safety alignment?

Key Contributions and Findings

1. Taxonomy of Safety Alignment Approaches (RQ1)

The authors propose a taxonomy of safety alignment methods based on three primary adaptation mechanisms:

• Data Adaptation: This approach prioritizes the creation of culturally grounded training data.
  • Findings: Native-language data construction outperforms translated English data. Methods include generating culturally specific datasets (e.g., CultureGuard, SEALGuard), using high-quality translation pipelines that preserve intent (e.g., LionGuard 2), and employing toxicity-preserving translation frameworks.
  • Limitation: Direct translation of training data often reduces performance and fails to capture local harm definitions.
• Objective Optimization: This strategy modifies training objectives to transfer safety behaviors without massive new datasets.
  • Findings: Techniques include Multilingual Reward Gap Optimization (MPO) to minimize reward gaps across languages, and Knowledge Distillation (though unfiltered distillation can increase jailbreak success). Reasoning-based approaches (e.g., MrGuard, ConsistentGuard) that generate explicit reasoning chains show improved generalization and robustness against code-switching attacks.
• Mechanistic Alignment: This involves direct modification of model internals.
  • Findings: Methods like Sparse Weight Editing (SWE) and layer-level activation analysis identify that safety capabilities are concentrated in sparse subsets of parameters. Transplanting these specific layers or weights from high-resource models to low-resource ones can improve safety without full retraining.

2. Safety Risks and Adversarial Behaviors (RQ2)

The review identifies specific vulnerabilities in multilingual LLMs:

• Cross-Lingual Jailbreaks: Safety alignment fails to generalize; translating harmful English prompts into LRLs can yield jailbreak success rates as high as 79% (compared to <15% in high-resource languages).
• Code-Switching Attacks: Mixing languages within a single prompt significantly reduces safety robustness, with bypass rates reaching ~67% on GPT-3.5. This is particularly critical in regions where code-switching is a normative communication style.
• Fine-Tuning as an Attack Vector: Adapting aligned models to new languages, even with benign data, can degrade previously established safety constraints.
• Culturally Specific Harms: Models often fail to recognize culturally rooted harms (e.g., toxic product recommendations in Hausa) or misinterpret composite harms during translation, leading to the propagation of stereotypes missed by English-centric benchmarks.

3. Evaluation Benchmarks (RQ3)

The landscape of safety benchmarks is shifting from direct translation to native, culturally grounded evaluation, though significant gaps remain:

• Global Benchmarks: Resources like XSafety and LinguaSafe are expanding to include multiple languages but often rely on translation or trans-creation.
• Regional Benchmarks: Southeast Asian and South Asian regions have seen the development of localized benchmarks (e.g., SEALSBench, IndicSafe).
• African Language Gap: Benchmarks for African languages are severely under-represented. Existing resources like LSR and Ubuntu-Guard are emerging but are insufficient compared to other multilingual ecosystems. The review notes a persistent lack of native-language safety benchmarks for African languages.

4. Factors Affecting Cross-Lingual Transfer (RQ4)

The review identifies several drivers of safety degradation in LRLs:

• Uneven Pre-training Coverage: Safety-relevant features are concentrated in high-resource regions of the latent space. Models perform poorly in languages with sparse pre-training data.
• Linguistic and Structural Differences: Morphologically rich languages and underrepresented scripts suffer from fragmented tokenization, weakening semantic representations and safety reasoning.
• Transfer Instability: Introducing new languages during fine-tuning can disrupt safety alignment.
• Representation Limitations: Safety transfer is hindered by the lack of shared multilingual safety neurons and the difficulty of aligning reasoning processes across diverse linguistic structures.

Significance and Limitations

Significance:
This paper provides the first systematic literature review dedicated to LLM safety alignment in African and low-resource languages. It synthesizes fragmented research into a coherent taxonomy and highlights a persistent "multilingual safety gap." The authors argue that future progress requires a shift away from English-centric assumptions toward culturally grounded benchmarks, participatory data collection, balanced multilingual pre-training, and scalable alignment methods that account for the specific linguistic and cultural contexts of low-resource communities.

Limitations:
The authors acknowledge that the review is limited to English-language studies from major academic databases, potentially missing locally published or recent research in native languages. Furthermore, grouping languages into broad regional categories may obscure important linguistic nuances. The paper does not propose new experimental methods but rather synthesizes existing findings to guide future research directions.