Article

Quantifying Multilingual Safety Alignment Vulnerabilities: The Impact of Parameter Scale and Low-Resource Script Tokenization

2026-08-31

Abstract excerpt

<title>Abstract</title> <p>Safety alignment in Large Language Models relies heavily on Reinforcement Learning from Human Feedback and system instruction formatting. However, existing safety datasets and automated red-teaming benchmarks remain predominantly English-centric, creating severe evaluation blind spots when models process regional low-resource languages and non-Latin scripts. This paper presents an empir...

Identifiers and source

Literature Corpus work: c767b7de-8e0a-5052-8547-e0d73ae3dbd6
DOI: 10.21203/rs.3.rs-10855619/v1

Open publication

Related research

Semantic proximity does not establish scientific evidence. Graph snapshot: 2026-09-29.

Click a neighbor to travelStep 1 · 12 closest

Quantifying Multilingual Safety Alignment Vulnerabilities: The Impact of Parameter Scale and Low-Resource Script TokenizationDOI 10.21203/rs.3.rs-10855619/v1Read

Harm-Conditioned Computational Friction as a Contrastive Diagnostic of Safety Representations in Language Models: A Proof-of-Concept and Control AuditSemantic proximity Read

Impact of Non-Standard Unicode Characters on Security and Comprehension in Large Language ModelsSemantic proximity Read

Investigating Reproducibility Challenges in LLM Bugfixing on the HumanEvalFix BenchmarkSemantic proximity Read

The Dice Roll Method: A StandardizedProtocol for Measuring Stochastic Bias in Large Language Model OutputSemantic proximity Read

Contamination Percolation in Multi-Agent LLM Systems: A Measurement Framework and BenchmarkSemantic proximity Read

AI-Induced Supply-Chain Compromise: A Systematic Review of Package Hallucinations and Slopsquatting AttacksSemantic proximity Read

Negative Preference Reduction in Large Language Model Unlearning: An Experimental ApproachSemantic proximity Read

Evaluating the Quality of Large Language Model-Generated Cybersecurity Advice in GRC SettingsSemantic proximity Read

The Invariant Energetic Cost of Constraint Persistence: A Self-Referential Framework for Measuring Alignment RobustnessSemantic proximity Read

SAMF: SAWANT (Structured Agentic Workflow for Alignment, Validation, and Negotiated Testing) for Reliable, Safe, and Verifiable LLM PromptingSemantic proximity Read

Harm-Conditioned Computational Friction as a Diagnostic of Alignment Robustness: A Critical Review and Evaluation FrameworkSemantic proximity Read

Finding Rare Racial Stereotypes in LLM Responses via Adversarial Anomaly DetectionSemantic proximity Read