Research map / Large language models / Instruction and fine-tuning
Toward Safe Quantization-Aware Fine-tuning: Understanding and Mitigating Safety Alignment Degradation
Yuning Yang, Guowei Peng, Xiurui Xie, Minrui Jiang, Shuang Liang, Guisong Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| quantize LoRA to 4-bit (QLoRA) | LoRA fine-tuning · SST5, Llama-3-8B-Instruct | Harmful Score: +8.2 Harmful Score, about 0.9 lower Finetune Accuracy | Hurt |
| quantize LoRA to 4-bit (QLoRA) | LoRA fine-tuning · Llama-3-8B-Instruct linear probe | probe accuracy drop: 1.73% -> 3.52% | Hurt |
| quantize LoRA to 4-bit (QLoRA) | LoRA fine-tuning · VISAGE safety metric | VISAGE drop: 6.34 -> 13.6 | Hurt |
| add safety-aware initialization and refinement (ExSQF) to QLoRA | QLoRA · SST5, avg over poison ratios | Harmful Score: 52.3 -> 14.4 | Improved |
| add safety matrix projection to LoftQ initialization | LoftQ · SST5, avg over poison ratios | Harmful Score: 29.7 -> 14.4 | Improved |
| initialize adapters from principal components of pre-trained weights (QPiSSA) instead of QLoRA | QLoRA · SST5, avg over poison ratios | Harmful Score: 52.3 -> 70.4 | Hurt |
| initialize adapters from quantization error (LoftQ) instead of QLoRA | QLoRA · SST5, avg over poison ratios | Harmful Score: 52.3 -> 29.7 | Improved |
| apply safety projection on output matrix O only versus Q, K, or V (ablation) | ExSQF refinement · SST5 | Harmful Score: 15.8 -> 12.6 | Improved |
| apply safety projection to all of Q/K/V/O instead of O only (ablation) | ExSQF refinement · SST5 | Harmful Score: 12.6 -> 15.8 | Hurt |
| transfer SafeLoRA and AsFT to quantized models (ablation) | full-precision safety methods · SST5 poison ratios | Harmful Score | Mixed |
| use ExSQF on GSM8K fine-tuning | ExSQF · GSM8K | Finetune Accuracy: 69.4 -> 75.0 | Improved |
| use ExSQF at high learning rates | ExSQF · SST5 learning rate sweep | Harmful Score | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 3 on most shared results.
- AsFT — AsFT: Anchoring Safety During LLM Fine-Tuning Within Narrow Safety Basin (AAAI 2026) · better on 3 of 4 shared results
- LoftQ — LoftQ: LoRA-Fine-Tuning-aware Quantization for Large Language Models (ICLR 2024) · better on 4 of 4 shared results
- SafeLoRA — Safe LoRA: The Silver Lining of Reducing Safety Risks when Finetuning Large Language Models (NeurIPS 2024) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Assessing Safety Risks and Quantization-aware Safety Patching for Quantized Large Language Models (ICML 2025 · similarity 0.94)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.91)
- Safe Delta: Consistently Preserving Safety when Fine-Tuning LLMs on Diverse Datasets (ICML 2025 · similarity 0.90)
- Booster: Tackling Harmful Fine-tuning for Large Language Models via Attenuating Harmful Perturbation (ICLR 2025 · similarity 0.90)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.89)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
