mlsanityData to 7 Oct 2026

Decoupling Safety into Orthogonal Subspace: Cost-Efficient and Performance-Preserving Alignment for Large Language Models

arXiv 2510.09004 ↗First posted 10 Oct 2025cs.CLcited by 0

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

also in 1 more
Abstract (arXiv)

Safety alignment is essential for building trustworthy artificial intelligence, yet it remains challenging to enhance model safety without degrading general performance. Current approaches require computationally expensive searches for the optimal proportion of safety-critical and general-purpose data to balance safety and general performance, incurring high costs with limited gains. In this work, we show that LoRA-based Refusal-training enables performance-preserving safety alignment even when trained solely on safety data, demonstrating that LoRA serves as cost-efficient, performance-preserving, and plug-and-play safety patches. Beyond empirical findings, we provide both theoretical and experimental evidence that LoRA effectively decouples safety into a low-rank subspace largely orthogonal to the model's intrinsic transformation space, ensuring that safety enhancements do not interfere with inherent capabilities.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Safety alignment and jailbreak attacks, one of 10 directions in LLM alignment / RLHF / preference optimisation. Builds on 48 of its 61 references that are in the map. First posted on arXiv 10 Oct 2025.

Builds on

48 of 61 references.

Closest ancestors · 12

Same direction, newest first

  1. 2025Lifelong Safety Alignment for Language Models26 May
  2. 2025MTSA: Multi-turn Safety Alignment for LLMs through Multi-round Red-teaming22 May
  3. 2025SaRO: Enhancing LLM Safety through Reasoning-based Alignment13 Apr
  4. 2025Safety Tax: Safety Alignment Makes Your Large Reasoning Models Less Reasonable1 Mar
  5. 2025X-Boundary: Establishing Exact Safety Boundary to Shield LLMs from Multi-Turn Jailbreaks without Compromising Usability14 Feb
  6. 2025STAIR: Improving Safety Alignment with Introspective Reasoning4 Feb
  7. 2024Data Advisor: Dynamic Data Curation for Safety Alignment of Large Language Models7 Oct
  8. 2024PKU-SafeRLHF: Towards Multi-Level Safety Alignment for LLMs with Human Preference20 Jun
  9. 2024Safety Alignment Should Be Made More Than Just a Few Tokens Deep10 Jun
  10. 2024Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications7 Feb
  11. 2023Safe RLHF: Safe Reinforcement Learning from Human Feedback19 Oct
  12. 2023BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Dataset10 Jul

Other references · 36

  1. 2023GPT-4 Technical Reportcited by 18,132
  2. 2022Training language models to follow instructions with human feedbackcited by 15,712
  3. 2017Proximal Policy Optimization Algorithmscited by 14,441
  4. 2024The Llama 3 Herd of Modelscited by 14,192
  5. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 7,916
  6. 2024Qwen2.5 Technical Reportcited by 4,713
  7. 2024GPT-4o System Cardcited by 4,467
  8. 2022Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedbackcited by 3,634
  9. 2023Jailbroken: How Does LLM Safety Training Fail?cited by 1,565
  10. 2024LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Modelscited by 1,442
  11. 2018Continual Lifelong Learning with Neural Networks: A Reviewcited by 1,313
  12. 2023LIMA: Less Is More for Alignmentcited by 1,080
24 more
  1. 2023Enhancing Chat Language Models by Scaling High-quality Instructional Conversationscited by 825
  2. 2023Safety-Tuned LLaMAs: Lessons From Improving the Safety of Large Language Models that Follow Instructionscited by 377
  3. 2023Orca: Progressive Learning from Complex Explanation Traces of GPT-4cited by 369
  4. 2024Improving Alignment and Robustness with Circuit Breakerscited by 302
  5. 2023Orthogonal Subspace Learning for Language Model Continual Learningcited by 296
  6. 2023MQuAKE: Assessing Knowledge Editing in Language Models via Multi-Hop Questionscited by 292
  7. 2024Deliberative Alignment: Reasoning Enables Safer Language Modelscited by 292
  8. 2024R-Judge: Benchmarking Safety Risk Awareness for LLM Agentscited by 268
  9. 2024SALAD-Bench: A Hierarchical and Comprehensive Safety Benchmark for Large Language Modelscited by 225
  10. 2023How Abilities in Large Language Models are Affected by Supervised Fine-tuning Data Compositioncited by 222
  11. 2024Rainbow Teaming: Open-Ended Generation of Diverse Adversarial Promptscited by 189
  12. 2023MART: Improving LLM Safety with Multi-round Automatic Red-Teamingcited by 161
  13. 2023Purple Llama CyberSecEval: A Secure Coding Benchmark for Language Modelscited by 130
  14. 2024ALERT: A Comprehensive Benchmark for Assessing Large Language Models' Safety through Red Teamingcited by 96
  15. 2024Detoxifying Large Language Models via Knowledge Editingcited by 86
  16. 2024Towards Lifelong Learning of Large Language Models: A Surveycited by 60
  17. 2025SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learningcited by 60
  18. 2024Analyzing and Reducing Catastrophic Forgetting in Parameter Efficient Tuningcited by 59
  19. 2023How Trustworthy are Open-Source LLMs? An Assessment under Malicious Demonstrations Shows their Vulnerabilitiescited by 33
  20. 2024Safety-Aware Fine-Tuning of Large Language Modelscited by 33
  21. 2024SEAS: Self-Evolving Adversarial Safety Optimization for Large Language Modelscited by 7
  22. 2023Continual Generalized Intent Discovery: Marching Towards Dynamic and Open-world Intent Recognitioncited by 5
  23. 2025TRIDENT: Enhancing Large Language Model Safety with Tri-Dimensional Diversified Red-Teaming Data Synthesiscited by 4
  24. 2024VersaTune: An Efficient Data Composition Framework for Training Multi-Capability LLMscited by 3

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026LSSF: Safety Alignment for Large Language Models through Low-Rank Safety Subspace Fusion
  2. 2025Safe Pruning LoRA: Robust Distance-Guided Pruning for Safety Alignment in Adaptation of LLMs
  3. 2025A Guardrail for Safety Preservation: When Safety-Sensitive Subspace Meets Harmful-Resistant Null-Space
  4. 2025LoRA is All You Need for Safety Alignment of Reasoning LLMs
  5. 2026Safety at One Shot: Patching Fine-Tuned LLMs with A Single Instance
  6. 2026Preventing Safety Drift in Large Language Models via Coupled Weight and Activation Constraints
  7. 2026Aligned but Fragile: Enhancing LLM Safety Robustness via Zeroth-Order Optimization
  8. 2026SafeAnchor: Preventing Cumulative Safety Erosion in Continual Domain Adaptation of Large Language Models