Research map / Large language models / Instruction and fine-tuning
From Parameter Dynamics to Risk Scoring: Quantifying Sample-Level Safety Degradation in LLM Fine-tuning
Xiao Wang, Yifei Zhang, YongKang Liu, Xiaocui Yang, Zihan Wang, Shi Feng, Daling Wang
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 |
|---|---|---|---|
| add module-wise normalization of sample-induced updates (ablation) | SQSD · CategoricalHarmfulQA, Dolly, Qwen3-8B | ASR difference S1-S5: 12.54 -> 68.72 | Improved |
| contrast safety and danger projections instead of danger only (ablation) | SQSD · CategoricalHarmfulQA, Dolly, Qwen3-8B | ASR difference S1-S5: 64.54 -> 68.72 | Improved |
| contrast safety and danger projections instead of safety only (ablation) | SQSD · CategoricalHarmfulQA, Dolly, Qwen3-8B | ASR difference S1-S5: 20.91 -> 68.72 | Improved |
| initialize at direction-sensitive parameter states (ablation) | SQSD · CategoricalHarmfulQA, Dolly, Qwen3-8B | ASR of S1: 38.36 -> 71.27 | Improved |
| use SQSD scores instead of baselines to rank samples | risk ranking methods · CategoricalHarmfulQA, 12 configurations | monotonic ASR decrease S1 to S5: 10/12 vs 0/6 or 1/6 for baselines | Improved |
| use SQSD scores instead of Reward Model to rank samples | risk ranking methods · CategoricalHarmfulQA, 12 configurations | average ASR difference S1-S5: 43.76% -> 49.86% | Improved |
| transfer SQSD risk scores from Qwen3-8B to Qwen3-14B | SQSD · CategoricalHarmfulQA | ASR S1 to S5: 55.09 -> 7.09 | Improved |
| transfer SQSD scores computed from LoRA gradients to full fine-tuning | SQSD · CategoricalHarmfulQA | ASR S1 to S5: 10.73 -> 2.55 | Improved |
| increase benign fine-tuning data scale from 3k to 50k Alpaca samples (ablation) | Qwen3-8B LoRA fine-tuning · Alpaca, Safety Score | Safety Score: 3.8 -> -0.3 | Hurt |
Closest papers by meaning
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.93)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.92)
- Safe Delta: Consistently Preserving Safety when Fine-Tuning LLMs on Diverse Datasets (ICML 2025 · similarity 0.91)
- Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons (NeurIPS 2025 · similarity 0.91)
- What Makes and Breaks Safety Fine-tuning? A Mechanistic Study (NeurIPS 2024 · similarity 0.91)
- Do as I do (Safely): Mitigating Task-Specific Fine-tuning Risks in Large Language Models (ICLR 2025 · similarity 0.90)
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.
