MEUV: Achieving Fine-Grained Capability Activation in Large Language Models via Mutually Exclusive Unlock Vectors
On the map: LLM alignment / RLHF / preference optimisation → Safety alignment and jailbreak attacks
also in 1 more
Abstract (arXiv)
Large language models (LLMs) enforce safety alignment to reliably refuse malicious requests, yet the same blanket safeguards also block legitimate uses in policing, defense, and other high-stakes settings. Earlier "refusal-direction" edits can bypass those layers, but they rely on a single vector that indiscriminately unlocks all hazardous topics, offering no semantic control. We introduce Mutually Exclusive Unlock Vectors (MEUV), a lightweight framework that factorizes the monolithic refusal direction into topic-aligned, nearly orthogonal vectors, each dedicated to one sensitive capability. MEUV is learned in a single epoch with a multi-task objective that blends a differential-ablation margin, cross-topic and orthogonality penalties, and several auxiliary terms. On bilingual malicious-prompt benchmarks, MEUV achieves an attack success rate of no less than 87% on Gemma-2-2B, LLaMA-3-8B, and Qwen-7B, yet cuts cross-topic leakage by up to 90% compared with the best single-direction baseline. Vectors trained in Chinese transfer almost unchanged to English (and vice versa), suggesting a language-agnostic refusal subspace. The results show that fine-grained, topic-level capability activation is achievable with minimal utility loss, paving the way for controlled LLMs deployment in security-sensitive domains.
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Safety alignment and jailbreak attacks, one of 10 directions in LLM alignment / RLHF / preference optimisation. Builds on 26 of its 28 references that are in the map. First posted on arXiv 4 Sep 2025.
Builds on
26 of 28 references.
Closest ancestors · 8
Same direction, newest first
- 2025The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independence24 Feb
- 2024Data Advisor: Dynamic Data Curation for Safety Alignment of Large Language Models7 Oct
- 2024SciSafeEval: A Comprehensive Benchmark for Safety Alignment of Large Language Models in Scientific Tasks2 Oct
- 2024Automatic Pseudo-Harmful Prompt Generation for Evaluating False Refusals in Large Language Models1 Sep
- 2024One-Shot Safety Alignment for Large Language Models via Optimal Dualization29 May
- 2024Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications7 Feb
- 2023AutoDAN: Interpretable Gradient-Based Adversarial Attacks on Large Language Models23 Oct
- 2023Universal and Transferable Adversarial Attacks on Aligned Language Models27 Jul
Other references · 18
- 2022Training language models to follow instructions with human feedbackcited by 15,971
- 2024The Llama 3 Herd of Modelscited by 14,715
- 2020Extracting Training Data from Large Language Modelscited by 2,189
- 2024Qwen2 Technical Reportcited by 2,120
- 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,998
- 2023Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injectioncited by 1,443
- 2022LaMDA: Language Models for Dialog Applicationscited by 1,220
- 2024Refusal in Language Models Is Mediated by a Single Directioncited by 1,032
- 2023C-Pack: Packed Resources For General Chinese Embeddingscited by 578
- 2023XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Modelscited by 505
- 2023Multi-step Jailbreaking Privacy Attacks on ChatGPTcited by 333
- 2023Large Language Model Alignment: A Surveycited by 243
6 more
- 2024Weak-to-Strong Jailbreaking on Large Language Modelscited by 107
- 2023MarioGPT: Open-Ended Text2Level Generation through Large Language Modelscited by 64
- 2024Revisiting Jailbreaking for Large Language Models: A Representation Engineering Perspectivecited by 48
- 2024Navigating the OverKill in Large Language Modelscited by 44
- 2024Jailbreaking Large Language Models Against Moderation Guardrails via Cipher Characterscited by 42
- 2024CPSDBench: A Large Language Model Evaluation Benchmark and Baseline for Chinese Public Security Domaincited by 2
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2024Uncovering Safety Risks of Large Language Models through Concept Activation Vector
- 2026Faithful Dual-constrained Erasure for Robust LLM Safety Alignment
- 2025From Threat to Tool: Leveraging Refusal-Aware Injection Attacks for Safety Alignment
- 2025Improving LLM Safety Alignment with Dual-Objective Optimization
- 2026RAS: Measuring LLM Safety Through Refusal Alignment
- 2025TRIDENT: Enhancing Large Language Model Safety with Tri-Dimensional Diversified Red-Teaming Data Synthesis
- 2024Recent Advances in Attack and Defense Approaches of Large Language Models
- 2023Latent Jailbreak: A Benchmark for Evaluating Text Safety and Output Robustness of Large Language Models