Research map / Large language models / Instruction and fine-tuning
SPQR: A Multi-Dimensional Benchmark for Safety Alignment under Benign Model Adaptation
Mohammed Talha Alam, Nada Saadi, Fahad Shamshad, Nils Lukas, Karthik Nandakumar, Fakhri Karray, Samuele Poppi
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 |
|---|---|---|---|
| fine-tune with LoRA instead of full fine-tuning (ablation) | SalUn · SD1.5 benign fine-tuning | robustness R: 0.726 -> 1.000 | Improved |
| fine-tune with LoRA instead of full fine-tuning (ablation) | ESD · SD1.5 benign fine-tuning | robustness R: 0.684 -> 0.950 | Improved |
| fine-tune with LoRA instead of full fine-tuning (ablation) | AdvUnlearn · SD1.5 benign fine-tuning | robustness R: 0.159 -> 0.087 | Hurt |
| fine-tune with LoRA instead of full fine-tuning (ablation) | SPM · SD1.5 benign fine-tuning | robustness R: 0.684 -> 0.571 | Hurt |
| fine-tune with LoRA instead of full fine-tuning (ablation) | UCE · SD1.5 benign fine-tuning | robustness R: 0.942 -> 0.819 | Hurt |
| fine-tune cross-attention only instead of full UNet (ablation) | RECE · SD1.5 benign fine-tuning | robustness R: 0.980 -> 0.942 | No clear effect |
| switch benign fine-tuning data from English general to multilingual | ESD · SD1.5 | robustness R: 0.684 -> 0.291 | Hurt |
| switch benign fine-tuning data from English general to Arabic | MACE · SD1.5 | robustness R: 0.657 -> 0.979 | Improved |
| evaluate on adversarial RAB prompts instead of ViSU | ESD · SD1.5 standard profile | robustness R: 0.684 -> 0.020 | Hurt |
| evaluate RECE built on UCE | RECE vs UCE · SD1.5 general | robustness R: 0.942 -> 0.980 | Improved |
| benign fine-tuning raises utility while safety drops | EraseDiff · SD1.5 | delta CLIP and delta FID: +0.111, -216.764 | Improved |
| use raw CLIPScore for prompt adherence normalization (ablation) | SPQR score · ranking stability | Spearman rho: 1.000 | No clear effect |
Closest papers by meaning
- Towards Resilient Safety-driven Unlearning for Diffusion Models against Downstream Fine-tuning (NeurIPS 2025 · similarity 0.90)
- T2ISafety: Benchmark for Assessing Fairness, Toxicity, and Privacy in Image Generation (CVPR 2025 · similarity 0.90)
- Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models? (CVPR 2025 · similarity 0.89)
- OVERT: A Benchmark for Over-Refusal Evaluation on Text-to-Image Models (NeurIPS 2025 · similarity 0.89)
- Prompting4Debugging: Red-Teaming Text-to-Image Diffusion Models by Finding Problematic Prompts (ICML 2024 · similarity 0.88)
- SafeVid: Toward Safety Aligned Video Large Multimodal Models (NeurIPS 2025 · similarity 0.88)
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.
