mlsanityData to 14 Oct 2026

LASA: Language-Agnostic Semantic Alignment at the Semantic Bottleneck for LLM Safety

arXiv 2604.12710 ↗First posted 13 Apr 2026cs.LG cs.AI cs.CLcited by 3

On the map: LLM alignment / RLHF / preference optimisation → Safety alignment and jailbreak attacks

also in 1 more
Abstract (arXiv)

Large language models (LLMs) often demonstrate strong safety performance in high-resource languages, yet exhibit severe vulnerabilities when queried in low-resource languages. We attribute this gap to a mismatch between language-agnostic semantic understanding ability and language-dominant safety alignment biased toward high-resource languages. Consistent with this hypothesis, we empirically identify the semantic bottleneck in LLMs, an intermediate layer in which the geometry of model representations is governed primarily by shared semantic content rather than language identity. Building on this observation, we propose Language-Agnostic Semantic Alignment (LASA), which anchors safety alignment directly in semantic bottlenecks. Experiments show that LASA substantially improves safety across all languages: average attack success rate (ASR) drops from 24.7% to 2.8% on LLaMA-3.1-8B-Instruct and remains around 3-4% across Qwen2.5 and Qwen3 Instruct models (7B-32B). Together, our analysis and method offer a representation-level perspective on LLM safety, suggesting that safety alignment requires anchoring safety understanding not in surface text, but in the model's language-agnostic semantic space.

Key facts. Cited by 3 papers in the map (data through 14 Oct 2026). Placed in Safety alignment and jailbreak attacks, one of 10 directions in LLM alignment / RLHF / preference optimisation. Builds on 36 of its 52 references that are in the map. First posted on arXiv 13 Apr 2026.

Builds on

36 of 52 references.

Closest ancestors · 7

Same direction, newest first

  1. 2025Do Methods to Jailbreak and Defend LLMs Generalize Across Languages?1 Nov
  2. 2025MPO: Multilingual Safety Alignment via Reward Gap Optimization22 May
  3. 2025AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak Defender13 Apr
  4. 2025The Hidden Space of Safety: Understanding Preference-Tuned LLMs in Multilingual context3 Apr
  5. 2025LLM Safety Alignment is Divergence Estimation in Disguise2 Feb
  6. 2024PKU-SafeRLHF: Towards Multi-Level Safety Alignment for LLMs with Human Preference20 Jun
  7. 2024The Language Barrier: Dissecting Safety Challenges of LLMs in Multilingual Contexts23 Jan

Other references · 29

  1. 2024The Llama 3 Herd of Modelscited by 14,715
  2. 2025Qwen3 Technical Reportcited by 8,820
  3. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,626
  4. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 8,065
  5. 2020Measuring Massive Multitask Language Understandingcited by 7,963
  6. 2024Qwen2.5 Technical Reportcited by 4,942
  7. 2025Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilitiescited by 3,796
  8. 2022BLOOM: A 176B-Parameter Open-Access Multilingual Language Modelcited by 1,984
  9. 2024HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusalcited by 1,429
  10. 2019Cross-lingual Language Model Pretrainingcited by 1,090
  11. 2024ORPO: Monolithic Preference Optimization without Reference Modelcited by 562
  12. 2022Language Models are Multilingual Chain-of-Thought Reasonerscited by 530
17 more
  1. 2024Contrastive Preference Optimization: Pushing the Boundaries of LLM Performance in Machine Translationcited by 394
  2. 2023Low-Resource Languages Jailbreak GPT-4cited by 309
  3. 2023Preference Ranking Optimization for Human Alignmentcited by 285
  4. 2024KTO: Model Alignment as Prospect Theoretic Optimizationcited by 264
  5. 2024Direct Preference Optimization with an Offsetcited by 122
  6. 2023Don't Trust ChatGPT when Your Question is not in English: A Study of Multilingual Abilities and Types of LLMscited by 104
  7. 2024Detoxifying Large Language Models via Knowledge Editingcited by 87
  8. 2025The State of Multilingual LLM Safety Research: From Measuring the Language Gap to Mitigating Itcited by 45
  9. 2025Refusal Direction is Universal Across Safety-Aligned Languagescited by 38
  10. 2024Enhancing Multilingual Capabilities of Large Language Models through Self-Distillation from Resource-Rich Languagescited by 36
  11. 2024RTP-LX: Can LLMs Evaluate Toxicity in Multilingual Scenarios?cited by 36
  12. 2024Code-Switching Red-Teaming: LLM Evaluation for Safety and Multilingual Understandingcited by 34
  13. 2023Improving In-context Learning of Multilingual Generative Language Models with Cross-lingual Alignmentcited by 33
  14. 2024HSF: Defending against Jailbreak Attacks with Hidden State Filteringcited by 23
  15. 2024Nudging: Inference-time Alignment of LLMs via Guided Decodingcited by 23
  16. 2025False Sense of Security: Why Probing-based Malicious Input Detection Fails to Generalizecited by 19
  17. 2025When Smiley Turns Hostile: Interpreting How Emojis Trigger LLMs' Toxicitycited by 2

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 3.

3 papers.

  1. 2026Multilingual Safety Alignment via Self-Distillationcited by 3
  2. 2026Low-Resource Safety Failures Are Action Failures, Not Representation Failurescited by 2
  3. 2026MASCRDM: Multi-Agent System for Compliance Risk Detection and Mitigation in Training Process of Large Language Modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026LLM Safety Alignment in Low-Resource Languages: A Systematic Literature Review
  2. 2025SaRO: Enhancing LLM Safety through Reasoning-based Alignment
  3. 2024Safety Alignment Should Be Made More Than Just a Few Tokens Deep
  4. 2025Safety Alignment Depth in Large Language Models: A Markov Chain Perspective
  5. 2024Superficial Safety Alignment Hypothesis
  6. 2026When Safety Speaks a Language: A Mechanistic Analysis of Safety-Language Identity Entanglement in LLMs
  7. 2024Towards Comprehensive Post Safety Alignment of Large Language Models via Safety Patching
  8. 2026Dialectics of Alignment: Harnessing Unsafe Knowledge for Dynamic Safety Routing