Why Do Safety Guardrails Degrade Across Languages?

TL;DR

This paper introduces a latent variable model to analyze safety degradation in large language models across languages, revealing nuanced factors influencing safety failures and enabling fairer cross-lingual safety evaluation.

Contribution

The authors develop a Multi-Group Item Response Theory framework that decouples safety factors, providing detailed insights into cross-lingual safety vulnerabilities of language models.

Findings

Safety is primarily unidimensional, with models refusing harm types through a shared mechanism.

22 configurations are more vulnerable in English than in low-resource languages.

High cross-lingual safety gap prompts cluster in physical harm categories and lower-resource languages.

Abstract

Large language models exhibit safety degradation in non-English languages. Standard evaluation relies on Jailbreak Success Rate (JSR), which confounds several safety-driving factors into one, obscuring the specific cause(s) of safety failure. We introduce a latent variable model, a Multi-Group Item Response Theory (IRT) framework, that decouples safety-driving factors such as language-agnostic safety robustness ($\theta$), intrinsic prompt hardness ($\beta$), global language processing difficulty ($\gamma$), and a prompt-specific cross-lingual safety gap ($\tau$). Using the MultiJail dataset, we evaluate the safety robustness of 61 model configurations across 5 closed-model families and 10 languages of varying resource, aggregating a dataset of 1.9 million rows. Exploratory Factor Analysis shows safety is primarily unidimensional: models refuse different harm types mainly through a shared mechanism. Contrary to the expected trend that safety degrades largely in low-resource languages, 22 model configurations are more vulnerable in English than in low-resource languages. Low-resource languages produce more uncertain responses (high entropy) than high-resource languages. Also, high-$\tau$ prompts cluster in physical harm categories like Theft and Weapons and lower-resource languages, trends validated through cross-dataset generalization. While global translation quality shows low correlation with $\tau$, severe mistranslations drive high-bias outliers, as validated by native speakers. Cultural and conceptual grounding mismatches also contribute to $\tau$. In predictive validation, the IRT framework achieves $\mathrm{AUC} = 0.940$, outperforming simpler baselines in predicting safe refusal of unsafe prompts. Our framework reveals concept-language vulnerabilities that aggregate metrics obscure, enabling fairer cross-lingual safety evaluation and targeted improvements in dataset construction.