You Snooze, You Lose: Automatic Safety Alignment Restoration through Neural Weight Translation
cs.CR
Submitted: 2026-05-06
Updated: 2026-08-27
Code: https://github.com/AI4LIFE-GROUP/med-safety
License: http://creativecommons.org/licenses/by/4.0/
The gist: 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
Terminology
Abstract
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.
Sources
- Concrete Problems in AI Safety
- SecureBreak -- A dataset towards safe and secure models
- XBreaking: Understanding how LLMs security alignment can be broken
- LoRA as Oracle
- Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback
- Constitutional AI: Harmlessness from AI Feedback
- Red-Teaming Large Language Models using Chain of Utterances for Safety-Alignment
- Unveiling the Basin-Like Loss Landscape in Large Language Models
- HuatuoGPT-o1, Towards Medical Complex Reasoning with LLMs
- Training Verifiers to Solve Math Word Problems
- Safe RLHF: Safe Reinforcement Learning from Human Feedback
- Safeguard Fine-Tuned LLMs Through Pre- and Post-Tuning Model Merging
- Omni-MATH: A Universal Olympiad Level Mathematic Benchmark For Large Language Models
- The Aloe Family Recipe for Open and Specialized Healthcare LLMs
- The Llama 3 Herd of Models
- Parameter-Efficient Fine-Tuning for Large Models: A Comprehensive Survey
- Measuring Mathematical Problem Solving With the MATH Dataset
- LoRA: Low-Rank Adaptation of Large Language Models
- UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-world Document Analysis
- TRIDENT: Benchmarking LLM Safety in Finance, Medicine, and Law
Related papers
- SoK: AI-Augmented Binary Reversing
- Relaxed Sender Anonymity for CBDC Interbank Settlement: A Zero-Knowledge Approach on Permissioned EVM
- Calibration-Family Overfit: Why Trusted Sabotage Monitors Don't Transfer Across Lineages
- Estimating the Social Cost of Corporate Data Breaches
- Efficient Fuzzy PSI under One-Sided Assumptions
- Sealing the Audit-Runtime Gap for LLM Skills