The Language Barrier: Dissecting Safety Challenges of LLMs in Multilingual Contexts
read the original abstract
As the influence of large language models (LLMs) spans across global communities, their safety challenges in multilingual settings become paramount for alignment research. This paper examines the variations in safety challenges faced by LLMs across different languages and discusses approaches to alleviating such concerns. By comparing how state-of-the-art LLMs respond to the same set of malicious prompts written in higher- vs. lower-resource languages, we observe that (1) LLMs tend to generate unsafe responses much more often when a malicious prompt is written in a lower-resource language, and (2) LLMs tend to generate more irrelevant responses to malicious prompts in lower-resource languages. To understand where the discrepancy can be attributed, we study the effect of instruction tuning with reinforcement learning from human feedback (RLHF) or supervised finetuning (SFT) on the HH-RLHF dataset. Surprisingly, while training with high-resource languages improves model alignment, training in lower-resource languages yields minimal improvement. This suggests that the bottleneck of cross-lingual alignment is rooted in the pretraining stage. Our findings highlight the challenges in cross-lingual LLM safety, and we hope they inform future research in this direction.
This paper has not been read by Pith yet.
Forward citations
Cited by 7 Pith papers
-
ROK-FORTRESS: Measuring the Effect of Geopolitical Transcreation for National Security and Public Safety
ROK-FORTRESS shows Korean-language prompts increase LLM safety suppression compared with English, while Korean geopolitical grounding often reduces that suppression, indicating translation-only evaluations miss langua...
-
BIASEDTALES-ML: A Multilingual Dataset for Analyzing Narrative Attribute Distributions in LLM-Generated Stories
BiasedTales-ML provides a parallel multilingual corpus of LLM-generated children's stories that reveals substantial cross-lingual differences in narrative attributes not captured by English-centric analyses.
-
Precise Shield: Explaining and Aligning VLLM Safety via Neuron-Level Guidance
Precise Shield identifies safety neurons in VLLMs via activation contrasts and aligns only them with gradient masking, boosting safety, preserving generalization, and enabling zero-shot cross-lingual and cross-modal transfer.
-
To Lie or Not to Lie? Investigating The Biased Spread of Global Lies by LLMs
LLMs propagate misinformation more in lower-resource languages and lower-HDI countries, with input safety classifiers and retrieval-augmented fact-checking showing cross-lingual and regional gaps.
-
Response-Based Knowledge Distillation for Multilingual Jailbreak Prevention Unwittingly Compromises Safety
Distilling safe refusal behavior from OpenAI o1-mini into Llama-3, Gemma-2, and Qwen3 models via response-based LoRA on multilingual jailbreak data increases jailbreak success rates on MultiJail by up to 16.6 points.
-
StanceNakba Shared Task: Actor and Topic-Aware Stance Detection in Public Discourse
Introduces StanceNakba 2026 shared task and 2,606-post dataset for Pro-Palestine/Pro-Israel/Neutral stance in English posts and Favor/Against/Neither stance in Arabic posts on normalization and refugees, with top Macr...
-
Safety is Contextual, LLM-Judges Are Not: Navigating the Rigid Priors of Evaluators
LLM safety judges resist adjusting evaluations when given contradictory context or new safety definitions, despite some ability to learn from new information.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.