mlsanityData to 7 Oct 2026

You Snooze, You Lose: Automatic Safety Alignment Restoration through Neural Weight Translation

arXiv 2605.04992 ↗First posted 6 May 2026cs.CRcited by 0

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

also in 1 more
Abstract (arXiv)

Public repositories now distribute thousands of specialized Low-Rank Adaptation (LoRA) modules, but a third-party adapter routinely arrives without the refusal behavior the same adapter would have had if it had been trained responsibly. Restoring it by fine-tuning on safety data induces the opposite failure: the domain expertise the adapter was published to provide degrades. A practitioner who downloads a finished adapter holds neither the domain corpus nor a safety corpus, so neither remedy is available. We propose Neural Weight Translation (NeWTral), which maps unsafe domain adapters to their safety-aligned counterparts entirely within parameter space while preserving their expertise. NeWTral is a non-linear translator pre-trained on unsafe-to-safe adapter pairs, with a Mixture of Experts router that blends a high-fidelity surgical translator and an aggressive alignment expert per tensor. Across four architectural families (Llama, Mistral, Qwen, Gemma) at scales up to 72B parameters and eight professional domains, NeWTral cuts the average Attack Success Rate from 70% to 13% while retaining 88% knowledge fidelity. The effect reproduces on JailbreakBench, on 12 of 14 adapters downloaded from a public model hub, and under automated jailbreak attacks. Curing is a single pass over the weights, needing no original training data and no retraining.

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 63 of its 119 references that are in the map. First posted on arXiv 6 May 2026.

Builds on

63 of 119 references.

Closest ancestors · 9

Same direction, newest first

  1. 2025NeuroStrike: Neuron-Level Attacks on Aligned LLMs15 Sep
  2. 2025Safe Pruning LoRA: Robust Distance-Guided Pruning for Safety Alignment in Adaptation of LLMs21 Jun
  3. 2025Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback22 Mar
  4. 2024SciSafeEval: A Comprehensive Benchmark for Safety Alignment of Large Language Models in Scientific Tasks2 Oct
  5. 2024PKU-SafeRLHF: Towards Multi-Level Safety Alignment for LLMs with Human Preference20 Jun
  6. 2023Removing RLHF Protections in GPT-4 via Fine-Tuning9 Nov
  7. 2023LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B31 Oct
  8. 2023Safe RLHF: Safe Reinforcement Learning from Human Feedback19 Oct
  9. 2023Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!5 Oct

Other references · 54

  1. 2020Language Models are Few-Shot Learnerscited by 32,199
  2. 2022Training language models to follow instructions with human feedbackcited by 15,712
  3. 2024The Llama 3 Herd of Modelscited by 14,192
  4. 2021LoRA: Low-Rank Adaptation of Large Language Modelscited by 14,084
  5. 2021Training Verifiers to Solve Math Word Problemscited by 9,743
  6. 2019Sentence-BERT: Sentence Embeddings using Siamese BERT-Networkscited by 9,049
  7. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 7,916
  8. 2019BERTScore: Evaluating Text Generation with BERTcited by 5,462
  9. 2022Flow Matching for Generative Modelingcited by 5,426
  10. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,203
  11. 2017Deep reinforcement learning from human preferencescited by 3,846
  12. 2022Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedbackcited by 3,634
38 more
  1. 2023Let's Verify Step by Stepcited by 3,598
  2. 2023QLoRA: Efficient Finetuning of Quantized LLMscited by 3,355
  3. 2022Constitutional AI: Harmlessness from AI Feedbackcited by 2,950
  4. 2025DAPO: An Open-Source LLM Reinforcement Learning System at Scalecited by 2,678
  5. 2020Learning to summarize from human feedbackcited by 2,562
  6. 2018Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answeringcited by 2,114
  7. 2024Qwen2 Technical Reportcited by 2,076
  8. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,928
  9. 2016Concrete Problems in AI Safetycited by 1,582
  10. 2023Jailbroken: How Does LLM Safety Training Fail?cited by 1,565
  11. 2024HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusalcited by 1,330
  12. 2023Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversationscited by 1,124
  13. 2022Red Teaming Language Models with Language Modelscited by 983
  14. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 960
  15. 2024Parameter-Efficient Fine-Tuning for Large Models: A Comprehensive Surveycited by 641
  16. 2015Generating Sentences from a Continuous Spacecited by 607
  17. 2021Datasets: A Community Library for Natural Language Processingcited by 526
  18. 2023Zephyr: Direct Distillation of LM Alignmentcited by 498
  19. 2024JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Modelscited by 494
  20. 2023Is ChatGPT a General-Purpose Natural Language Processing Task Solver?cited by 486
  21. 2023XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Modelscited by 477
  22. 2024LoRA Learns Less and Forgets Lesscited by 313
  23. 2024HuatuoGPT-o1, Towards Medical Complex Reasoning with LLMscited by 257
  24. 2023Red-Teaming Large Language Models using Chain of Utterances for Safety-Alignmentcited by 213
  25. 2024Orca-Math: Unlocking the potential of SLMs in Grade School Mathcited by 151
  26. 2023Large Language Models in Law: A Surveycited by 128
  27. 2024Safe LoRA: the Silver Lining of Reducing Safety Risks when Fine-tuning Large Language Modelscited by 122
  28. 2024Bag of Tricks: Benchmarking of Jailbreak Attacks on LLMscited by 47
  29. 2023Deficiency of Large Language Models in Finance: An Empirical Examination of Hallucinationcited by 41
  30. 2024UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-world Document Analysiscited by 40
  31. 2025SaLoRA: Safety-Alignment Preserved Low-Rank Adaptationcited by 39
  32. 2025Limitations of Large Language Models in Clinical Problem-Solving Arising from Inflexible Reasoningcited by 29
  33. 2024Improving Physics Reasoning in Large Language Models Using Mixture of Refinement Agentscited by 22
  34. 2025Unveiling the Basin-Like Loss Landscape in Large Language Modelscited by 13
  35. 2024Safeguard Fine-Tuned LLMs Through Pre- and Post-Tuning Model Mergingcited by 11
  36. 2025The Aloe Family Recipe for Open and Specialized Healthcare LLMscited by 7
  37. 2026Understanding and Preserving Safety in Fine-Tuned LLMscited by 7
  38. 2025GateBreaker: Gate-Guided Attacks on Mixture-of-Expert LLMscited by 6

50 most cited of 54 All 54 (JSON)

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025Decoupling Safety into Orthogonal Subspace: Cost-Efficient and Performance-Preserving Alignment for Large Language Models
  2. 2026SafeAnchor: Preventing Cumulative Safety Erosion in Continual Domain Adaptation of Large Language Models
  3. 2026Dialectics of Alignment: Harnessing Unsafe Knowledge for Dynamic Safety Routing
  4. 2026Safety-Utility Conflicts Are Not Global: Surgical Alignment via Head-Level Diagnosis
  5. 2026Safe Transformer: An Explicit Safety Bit For Interpretable And Controllable Alignment
  6. 2024NLSR: Neuron-Level Safety Realignment of Large Language Models Against Harmful Fine-Tuning
  7. 2026CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment
  8. 2026Safety at One Shot: Patching Fine-Tuned LLMs with A Single Instance