The study investigates how safety alignment in large language models, trained mainly in English, transfers to other languages. While models show near-perfect harmfulness detection (AUROC > 0.98) using unrelated harmless prompts (easy negatives), performance drops sharply in low‑resource languages when using surface‑similar benign prompts (hard negatives). This degradation persists across multiple languages and models, indicating that easy‑negative evaluation alone cannot confirm cross‑lingual harmfulness representation quality.
By Paras Balani, Subhrakanta Panda
arXiv:2606. 01196v1 Announce Type: cross Abstract: Safety alignment learned in high-resource languages transfers poorly to low-resource languages.
By Rashad Aziz, Ikhlasul Akmal Hanif, Fajri Koto
The paper introduces XIH-Bench, a multilingual benchmark for evaluating instruction hierarchy (IH) compliance in large language models across six languages, four domains, and three IH settings. It finds that IH compliance varies with language, showing an asymmetry where a language can strengthen compliance in a higher-priority role but disrupt it in a lower-priority role, and that cross-language conflicts yield higher compliance than same-language conflicts—a phenomenon called the Language Boundary Effect. The study also highlights that language specialization can make lower-priority instructions in favored languages harder to override, raising potential reliability and security concerns.
By Jiwon Moon, Yerin Hwang, Kyomin Jung
arXiv:2607. 10112v1 Announce Type: cross Abstract: Safety alignment in large language models remains brittle across languages: prompts reliably refused in English can elicit harmful compliance in non-English and low-resource settings.
By Chigozirim Ifebi, Brent Kong, Ayushi Mehrotra
Safety alignment in large language models (LLMs) is largely developed in English, assuming these safeguards generalize across multilingual settings. However, this assumption remains underexplored and exposes a vulnerability in low-resource languages.
The paper investigates how large language models (LLMs) can be bypassed by translating harmful requests from English into low‑resource languages. By testing ten popular LLMs across 79 languages, the authors find that safe English completions often fail to remain safe after translation, with bypass rates between 21.9% and 44.4%. The risk is unevenly distributed across language families and varies significantly between models, making it hard to rank models by safety consistently.
By Jalil Alva, David Harper, Abigail E. Thornton, Rafael S. Villalobos Alva, Leonardo Kenji Minemura Suazo, Konstantinos Karageorgos, Diego Velazquez, Sergio A. Serrano, Marvin Vargas Flores, Aaron Schliem, Fernando F. Migone
The paper investigates how narrative wrapping can trick safety‑aligned language models into refusing a harmful request only when it is presented directly, but answering it when embedded in a role‑play or narrative context. It introduces GUISE, a multilingual benchmark that tests this vulnerability across English, modern Chinese, and Classical Chinese, and shows high attack success rates (up to 95.7%). The authors propose AXIS, a defense that aligns harmful‑request representations with the refusal direction and trains the model to refuse outright, achieving superior safety and usability scores on several large language models.
By Zhankai Ye, Yanning Wang, Yukai Jin, Bo Mei, Fangyi Li, Wei Wang, Shangqian Gao, Xin Liu
arXiv:2609.00760v1 Announce Type: new
Abstract: Large language models (LLMs) are increasingly trained to decline queries that fall outside their knowledge (knowledge-based refusal, KR) or violate saf...
By Yuri Son, Seunghee Kim, Hyuhng Joon Kim, Taeuk Kim
arXiv:2608. 11583v1 Announce Type: new Abstract: Safety alignment in large language models is often treated as a distributed property of the entire network, yet its practical brittleness suggests that refusal behavior may be concentrated in a smaller set of parameters.
By Mingyu Zong, Sampad Mohanty, Bhaskar Krishnamachari
arXiv:2608.29378v1 Announce Type: cross
Abstract: Arabic large language models must refuse harmful prompts without over-refusing benign or sensitive prompts, yet a single refusal rate hides this trad...
By Mohamad Zbib, Ammar Mohanna
The paper introduces Latent Space Refusal Anchoring (LSR‑Anchoring), a training‑free technique that extracts a refusal direction from English prompts and applies it to the residual stream of instruction‑tuned models at inference time. The primary variant, Mean‑Activation Steering (MAS), works across several architectures (Llama‑3‑8B, Llama‑3.1‑70B, Mistral‑7B‑Instruct, Qwen2.5‑7B), restoring safety for low‑resource African languages with minimal performance loss, while a refined SAE‑Derived Steering (SDS) further reduces KL divergence without degrading legitimate prompt performance. The method shows positive transfer for Yoruba, Igbo, Igala, and Hausa, but fails for Arabic, suggesting a geometric mismatch rather than a data scarcity issue.
By Godwin Abuh Faruna
The study investigates whether independently trained language‑model societies share a common packet language and how inherited interface states affect learning. A comprehensive audit of 30 pairwise interactions among six restricted societies shows that only one pair is fully interoperable, another is partially compatible, and the remaining 26 pairs fail across all alignment levels. Further experiments reveal that a globally trained communication interface can act as a severe negative‑transfer prior, but inherited interfaces never outperform fresh‑interface controls by the preregistered margin.
By Narcis Marincat