This work proposes aligning cross-lingual thoughts and responses (ACTR), a framework that improves multilingual safety alignment by strengthening the use of existing safety reasoning, and devise neuron-selective consistency optimization (NSCO), which uses a frozen judge model to reward agreement between the safety categories of reasoning traces and responses.
Abstract
Ensuring the safety of reasoning large language models (LLMs) across languages is essential for their reliable deployment. However, when exposed to jailbreak attacks in non-high-resource languages, these models may generate unsafe responses even when their reasoning traces identify safety risks. To address this issue, we propose aligning cross-lingual thoughts and responses (ACTR), a framework that improves multilingual safety alignment by strengthening the use of existing safety reasoning. Specifically, we first present the think gap score (TGS) to compare the normalized contributions of reasoning traces to attention outputs during response generation across languages, and use reasoning- trace substitution to measure the cross-lingual safety gap. Next, using a corpus of jailbreak queries, we assess neuron importance through changes in response representations caused by neuron masking and compare the high-importance neuron sets obtained with reasoning enabled and disabled to identify safety think neurons that support the use of safety reasoning. Finally, we devise neuron-selective consistency optimization (NSCO), which uses a frozen judge model to reward agreement between the safety categories of reasoning traces and responses while updating only the parameters associated with the selected neurons, requiring no human-annotated responses or preference data. Across two reasoning models, ACTR achieves lower average attack success rates than the evaluated state-of-the-art methods on AdvBench-X and MultiJail, with safety gains extending to unseen languages, while preserving or improving average performance on multilingual knowledge and mathematical reasoning tasks and limiting false refusals of benign requests. Warning: this paper contains examples with unsafe content.
This work introduces a reasoning-guided translation framework for constructing multilingual evaluations designed to preserve the structure and intended meaning of diverse benchmark tasks, and develops a source-aligned evaluation collection spanning seven safety, conversational, and mathematical reasoning resources acro...
Zahraa Al Sahili, I. Patras, Matthew Purver· 0 citations
IndicSafeEval, a persuasion-based jailbreak evaluation framework for Indian languages, is introduced and it is observed that different risk categories exhibit different levels of vulnerability, with some types of harmful content being significantly more susceptible to persuasion-based jailbreaks than others.
Saikat Mondal, Mamta Mamta, Deeksha Varshney et al.· 0 citations
This paper presents a methodology for building safety evaluation datasets that comprehensively cover the full spectrum of sensitive topics relevant to LLM safety, and releases a public repository containing the list of categorized Italian Wikipedia pages, the automatically generated prompts, and the standard prompt tem...
Elena Marafatto, R. Navigli· International Conference on...· 0 citations
SSRFT shows substantially greater robustness to prefilling attacks and better generalization to unseen jailbreak domains, while reducing over-refusal on benign queries and preserving the model's general capabilities, establishing safe-role internalization as an effective alternative to refusal-centric safety alignment.
Jin-Hao Pang, Jitai Hao, Qiang Huang et al.· 0 citations
ReFrame is a training-free multimodal input reframing framework where two agents share a lightweight locally deployed MLLM: the evidence-generation agent constructs complementary risk and utility evidence, and the rewrite-and-routing agent converts it into a safe proxy prompt and image-routing decision before calling t...
Wenzheng Jiang, Xuan-Kun Rong, Yuan-Zhao Zhai et al.· 0 citations