Research map / Large language models / Instruction and fine-tuning
A Guardrail for Safety Preservation: When Safety-Sensitive Subspace Meets Harmful-Resistant Null-Space
Bingjie Zhang, Yibo Yang, Renzhe, Dandan Guo, Jindong Gu, Philip Torr, Bernard Ghanem
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 null space projector (ablation) | GuardSpace (Llama-2-7B-Chat) · GSM8K fine-tuning | HS (lower better): 52.00 -> 3.60 | Improved |
| add null space projector (ablation) | GuardSpace (Llama-2-7B-Chat) · GSM8K fine-tuning | FA: 28.60 -> 28.00 | No clear effect |
| add safety-irrelevant subspace initialization (ablation) | GuardSpace (Llama-2-7B-Chat) · GSM8K fine-tuning | HS (lower better): 5.20 -> 3.60 | Improved |
| add safety-irrelevant subspace initialization (ablation) | GuardSpace (Llama-2-7B-Chat) · GSM8K fine-tuning | FA: 26.20 -> 28.00 | Improved |
| GuardSpace vs AsFT on Llama-2-7B-Chat | GuardSpace · SST-2, AGNEWS, GSM8K average | HS (lower better): 8.13 -> 2.40 | Improved |
| GuardSpace vs AsFT on Llama-2-7B-Chat | GuardSpace · SST-2, AGNEWS, GSM8K average | FA: 67.87 -> 69.75 | Improved |
| GuardSpace vs LoRA with 20% unsafe data | GuardSpace · GSM8K fine-tuning, 20% unsafe data | HS (lower better): 60.00 -> 2.40 | Improved |
| GuardSpace vs LoRA on Qwen-2-7B-Instruct | GuardSpace · GSM8K fine-tuning | FA: 66.40 -> 65.40 (HS 30.00 -> 3.20) | Mixed |
| GuardSpace vs LoRA on Gemma-2-9B-IT | GuardSpace · GSM8K fine-tuning | FA: 69.80 -> 70.20 (HS 50.00 -> 2.80) | Mixed |
| use more than 1024 trailing components (large adapter rank r=1024) (ablation) | GuardSpace · GSM8K | HS (lower better) | Hurt |
| increase number of harmful prompts for projector from 120 to >=220 (ablation) | GuardSpace · GSM8K | HS (lower better): 24.8 -> 4-6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AsFT — AsFT: Anchoring Safety During LLM Fine-Tuning Within Narrow Safety Basin (AAAI 2026) · better on 9 of 9 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
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.90)
- Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements (ICLR 2025 · similarity 0.90)
- Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons (NeurIPS 2025 · similarity 0.90)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.89)
- Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications (ICML 2024 · similarity 0.89)
- The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Analysis of Orthogonal Safety Directions (ICML 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.
