The Heterogeneous Safety Impacts of Benign Multilingual Fine-Tuning

• URL: http://arxiv.org/abs/2606.28843v1
• Date: Sat, 27 Jun 2026 10:05:36 GMT
• Title: The Heterogeneous Safety Impacts of Benign Multilingual Fine-Tuning
• Abstract summary: Fine-tuning a large language model can increase its tendency to respond to unsafe adversarial prompts.<n>We present the first comprehensive empirical study of this phenomenon in multilingual settings by fine-tuning Llama-3.2, Qwen3, and Gemma-3 models.<n>We find that safety outcomes are highly sensitive to both the choice of fine-tuning language and the evaluation language.
• Score: 13.578131927232292
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Fine-tuning a large language model is a ubiquitous method for enhancing its capability on a specific downstream task. However, prior work has shown that this increase in capability comes with a cost: it can increase a model's tendency to respond to unsafe adversarial prompts, even when fine-tuning with non-adversarial data. We present the first comprehensive empirical study of this phenomenon in multilingual settings by fine-tuning Llama-3.2, Qwen3, and Gemma-3 models using benign data translated across nine languages. We find that safety outcomes are highly sensitive to both the choice of fine-tuning language and the evaluation language, with adversarial compliance rates increasing four-fold in some settings. Multilingual safety drift is decoupled from general capability metrics, and occurs heterogeneously across languages and models. Fine-tuning in non-English languages often induces smaller internal representational drifts than English, but these shifts lead models to default to either exaggerated compliance or refusal. As such, assessing fine-tuning impacts solely in English provides inadequate assurance for deployment. To facilitate further research into these cross-lingual safety blind spots, we release the Multilingual-Benign-Tune dataset and the SORRY-Bench-Multilingual evaluation suite.

Related papers

• Inference-Time Steering for Cross-Lingual Factual Consistency in LLMs [51.56484100374058]
  Large Language Models (LLMs) demonstrate remarkable multilingual fluency, but their internal knowledge representations remain disproportionately biased toward high-resource languages.<n>We investigate whether these biases can be mitigated at inference time, forcing an English-prompted model to answer as if it were queried in target languages.<n>Experiments on Gemma 3 12B Instruct reveal persona prompting to be the strongest overall intervention, balancing efficacy, safety, and out-of-domain generalization.
  arXiv  Detail & Related papers  (2026-07-21T16:15:05Z)
• Sycophancy as a Multilingual Alignment Failure: How Safety Degrades Across Languages, Topics, and Models [3.2555924024638667]
  Safety-aligned large language models often exhibit sycophancy, which is the tendency to affirm users' opinions regardless of factual accuracy.<n>We present the first large-scale, multi-model evaluation of cross-lingual sycophancy, benchmarking textbfsix instruction-tuned models across textbf38 languages and textbf33 topic categories.
  arXiv  Detail & Related papers  (2026-06-07T04:50:40Z)
• Multilingual Safety Alignment via Self-Distillation [17.94152626632751]
  Large language models (LLMs) exhibit severe multilingual safety misalignment.<n>We propose a cross-lingual safeguard transfer framework named Multilingual Self-Distillation (MSD)<n>Our framework is flexible and can be integrated with different self-distillation strategies.
  arXiv  Detail & Related papers  (2026-05-03T14:22:49Z)
• Multilingual Refusal Alignment for Safer Large Language Models [53.64286756804503]
  We investigate whether single-language alignment transfers cross-lingually, how language consistency is preserved during training, and the resulting trade-offs with general knowledge capabilities.<n>We introduce RefusEU, a novel refusal alignment dataset covering 12 European languages, including a dedicated test set for evaluating current state-of-the-art models.<n>Our controlled Direct Preference Optimization (DPO) experiments provide two key insights: aligning models exclusively in English is insufficient to ensure cross-lingual safety, even for the same harm categories, whereas training on multilingual datasets can improve safety without degrading general performance, as measured by the Global MMLU benchmark.
  arXiv  Detail & Related papers  (2026-04-24T09:29:14Z)
• Precise Shield: Explaining and Aligning VLLM Safety via Neuron-Level Guidance [64.18179414483966]
  In real-world deployments, Vision-Language Large Models (VLLMs) face critical challenges from multilingual and multimodal composite attacks.<n>We propose Precise Shield, a two-stage framework that first identifies safety neurons by contrasting activation patterns between harmful and benign inputs.<n>This strategy substantially improves safety while preserving multilingual and multimodal generalization.
  arXiv  Detail & Related papers  (2026-04-10T02:42:52Z)
• Parallel Universes, Parallel Languages: A Comprehensive Study on LLM-based Multilingual Counterfactual Example Generation [49.2073409243885]
  Large language models (LLMs) excel at generating English counterfactuals and demonstrate multilingual proficiency.<n>We conduct automatic evaluations on both directly generated counterfactuals in the target languages and those derived via English translation across six languages.<n>We identify and categorize four main types of errors that consistently appear in the generated counterfactuals across languages.
  arXiv  Detail & Related papers  (2026-01-01T08:53:49Z)
• Conditions for Catastrophic Forgetting in Multilingual Translation [24.10629800866219]
  We identify the conditions that trigger catastrophic forgetting in multilingual fine-tuning.<n>We show that the relative scale between model and data size is a primary determinant of forgetting.<n>We also show that cross-lingual alignment can mitigate forgetting while also facilitating positive transfer to unseen target languages.
  arXiv  Detail & Related papers  (2025-10-22T12:54:00Z)
• MrGuard: A Multilingual Reasoning Guardrail for Universal LLM Safety [56.77103365251923]
  Large Language Models (LLMs) are susceptible to adversarial attacks such as jailbreaking.<n>This vulnerability is exacerbated in multilingual settings, where multilingual safety-aligned data is often limited.<n>We introduce a multilingual guardrail with reasoning for prompt classification.
  arXiv  Detail & Related papers  (2025-04-21T17:15:06Z)
• Vicinal Risk Minimization for Few-Shot Cross-lingual Transfer in Abusive Language Detection [19.399281609371258]
  Cross-lingual transfer learning from high-resource to medium and low-resource languages has shown encouraging results. We resort to data augmentation and continual pre-training for domain adaptation to improve cross-lingual abusive language detection.
  arXiv  Detail & Related papers  (2023-11-03T16:51:07Z)
• On Negative Interference in Multilingual Models: Findings and A Meta-Learning Treatment [59.995385574274785]
  We show that, contrary to previous belief, negative interference also impacts low-resource languages. We present a meta-learning algorithm that obtains better cross-lingual transferability and alleviates negative interference.
  arXiv  Detail & Related papers  (2020-10-06T20:48:58Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.