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.
arXiv:2608. 14626v1 Announce Type: cross Abstract: Large Language Models (LLMs) have achieved substantial progress in safety alignment, yet their safety guarantees remain significantly weaker in low-resource and multilingual settings than in high-resource languages.
The paper introduces MMSAFE, a multi-layer framework designed to identify safety-degrading data in multilingual large language models. It shows that safety signals are distributed across multiple layers and only partially shared across languages, unlike the single-layer assumption used in monolingual settings. Experiments demonstrate that MMSAFE reduces harmful-response rates by 60% compared to random filtering and outperforms the best single-layer baseline across various models, languages, and safety benchmarks.
arXiv:2606. 08451v1 Announce Type: cross Abstract: Safety-aligned large language models often exhibit sycophancy, which is the tendency to affirm users' opinions regardless of factual accuracy.
arXiv:2608. 09095v1 Announce Type: new Abstract: Uncovering the internal mechanisms underlying the safety capabilities of large language models (LLMs) is crucial for developing trustworthy artificial intelligence.
arXiv:2608.29936v1 Announce Type: cross Abstract: Safety alignment of large language models (LLMs) degrades across languages, yet the internal mechanism driving this asymmetry remains poorly understo...
The paper titled "Safety Alignment Illusion: The Cross-Lingual Safety Gap in LLMs" highlights that current safety alignment training for large language models is predominantly English-centric, leading to failures in non‑English languages. It introduces INCLUDE, a multilingual benchmark with 2,604 prompts in six languages (English, Hindi, Bengali, Marathi, Tamil, and Hinglish) to measure Indian‑centric socio‑cultural biases. Evaluation of ten open‑ and closed‑source LLMs shows that Bengali models exhibit the highest bias scores among open‑source models, while English shows the lowest bias in open‑source but the highest in closed‑source models.
arXiv:2606. 28843v1 Announce Type: cross Abstract: Fine-tuning a large language model is a ubiquitous method for enhancing its capability on a specific downstream task.
IndicSafeEval is a new evaluation framework that tests the safety robustness of large language models against persuasion-based jailbreak attacks in Indian languages. The benchmark covers ten safety-critical content categories, six persuasive strategies, and four languages—Hindi, Bengali, Marathi, and Punjabi—producing 7,200 adversarial prompts. Experiments show that model safety varies significantly across languages, prompt styles, and risk categories, highlighting gaps in current English-centric safety assessments.
arXiv:2605. 17173v2 Announce Type: replace-cross Abstract: Large language models exhibit safety degradation in non-English languages.
IndicSafeEval is a new evaluation framework that tests the safety robustness of large language models against persuasion-based jailbreak attacks in Indian languages. The benchmark includes 7,200 adversarial prompts covering ten safety-critical content categories, six persuasive strategies, and four languages (Hindi, Bengali, Marathi, Punjabi). Experiments show that model safety varies significantly across languages, prompt styles, and risk categories, revealing that current English-centric safety evaluations miss important multilingual vulnerabilities.
arXiv:2603.04410v3 Announce Type: replace-cross Abstract: While different stakeholders are trying to leverage Arabic Language Models (ALMs), safety alignment in ALMs remains largely underexplored, hi...
Multilingual safety evaluation of large language models (LLMs) has predominantly relied on direct translation (DT) of English benchmarks into target languages - an approach that converts surface-level linguistic form while failing to reflect the cultural context embedded in threat scenarios, social norms, and legal frameworks. We construct paired DT and culturally-adapted (CA) datasets via 1:1 seed matching for four languages - Korean (KO), Japanese (JA), Thai (TH), and Khmer (KM) - and compare Attack Success Rate (ASR) and Cultural Realism scores across four open-source LLM.