aiqbench analysis
This research addresses the critical challenge of maintaining safety alignment in multilingual large language models by revealing that harmful signals are often distributed across multiple neural layers rather than a single one. The proposed MMSAFE framework leverages this multi-layer insight to filter safety-degrading data more effectively, significantly reducing harmful responses compared to existing single-layer methods. For practitioners building global AI systems, this work underscores the necessity of cross-lingual architectural analysis to prevent silent degradation of model safety.
Multilingual Safety Signals Are Multi-Layered: Filtering Safety-Degrading Data for Safer LLMs
arXiv:2609.22144v1 Announce Type: new Abstract: Preserving safety alignment during large language models fine-tuning is critical, however, recent studies have demonstrated that even benign fine-tuning data may contain safety-degrading samples that silently undermine safety alignment. Existing approaches typically identify such samples using representations from a single safety-sensitive layer. While this assumption has shown effectiveness in monolingual settings, its validity for multilingual models remains unclear due to potential cross-lingual differences in representation patterns. Through a cross-lingual analysis, we show that sensitive layers are only partially shared across languages, with safety-relevant signals often distributed across multiple layers. Motivated by these observations, we propose MMSAFE, a multi-layer framework for multilingual safety-degrading data identification that captures both shared and language-specific safety signals. Extensive experiments across multiple models, languages, and safety benchmarks demonstrate that MMSAFE reduces the average harmful-response ratio by 60% compared with random filtering and achieves stronger average performance than the strongest single-layer baseline, demonstrating the effectiveness of multi-layer modeling for robust multilingual safety alignment.
Abstract:Preserving safety alignment during large language models fine-tuning is critical, however, recent studies have demonstrated that even benign fine-tuning data may contain safety-degrading samples that silently undermine safety alignment. Existing approaches typically identify such samples using representations from a single safety-sensitive layer. While this assumption has shown effectiveness in monolingual settings, its validity for multilingual models remains unclear due to potential cross-lingual differences in representation patterns. Through a cross-lingual analysis, we show that sensitive layers are only partially shared across languages, with safety-relevant signals often distributed across multiple layers. Motivated by these observations, we propose MMSAFE, a multi-layer framework for multilingual safety-degrading data identification that captures both shared and language-specific safety signals. Extensive experiments across multiple models, languages, and safety benchmarks demonstrate that MMSAFE reduces the average harmful-response ratio by 60% compared with random filtering and achieves stronger average performance than the strongest single-layer baseline, demonstrating the effectiveness of multi-layer modeling for robust multilingual safety alignment.
Accepted to EMNLP 2026 Main Conference. 16 pages, 11 figures
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2609.22144 [cs.CL]
(or arXiv:2609.22144v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2609.22144 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 265 words
Related coverage
J++ Lens: Jacobian Filtering Enables More Faithful Workspace Lenses
LessWrong · Oct 8, 2026
Introducing Instella-MoE: A State-of-the-Art Fully Open Mixture-of-Experts Language Model
AMD · Jul 24, 2026
Building trust in AI agents: A production-ready evaluation framework with Amazon Bedrock AgentCore
AWS · Oct 16, 2026
Research Note: Filtering Subversion-Relevant Information From Pretraining Data Is Feasible
LessWrong · Oct 5, 2026