Research map / Adversarial robustness and security / Attacks on text and graphs
EASE: Practical and Efficient Safety Alignment for Small Language Models
Haonan Shi, Guoli Wang, Tu Ouyang, An 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 safety reasoning (EASE) vs refusal training | Qwen2.5-1.5B-Instruct · WildJailbreak | ASR (%): 23.26% -> 6.90% | Improved |
| EASE vs deliberative alignment on safety | Qwen2.5-3B-Instruct · WildJailbreak | ASR (%): 11.20% -> 5.35% | Improved |
| selective reasoning (EASE) vs always-reason deliberative alignment on token cost | Qwen2.5-1.5B-Instruct · HellaSwag | generated tokens: 317.0 -> 31.0 | Improved |
| add reasoning boundary calibration (ablation) | EASE · HellaSwag | avg tokens: 272 -> 31 | Improved |
| add reasoning boundary calibration (ablation) | EASE · WildJailbreak | ASR (%): 6.5% -> 6.9% | Mixed |
| select vulnerable regions targeted instead of random (ablation) | EASE · WildJailbreak | ASR (%): 10.2% -> 6.9% | Improved |
| EASE vs refusal training on general ability | Qwen2.5-3B-Instruct · MMLU | accuracy (%): 61.91% -> 63.95% | Improved |
| EASE vs original Instruct on general ability | Qwen2.5-1.5B-Instruct · MMLU | accuracy (%): 58.08% -> 57.71% | No clear effect |
| use smaller LRM teacher (DeepSeek-Qwen-14B-Distill) instead of larger teacher (ablation) | EASE student distillation · WildJailbreak subset (250 prompts) | safety rate | Improved |
| use LRM instead of LLM as safety reasoning teacher (ablation) | EASE student distillation · WildJailbreak subset (250 prompts) | safety rate | Improved |
| use DeepSeek-Qwen-7B-Distill as a teacher (too small) (ablation) | EASE student distillation · WildJailbreak subset (250 prompts) | safety rate | Hurt |
Closest papers by meaning
- STAIR: Improving Safety Alignment with Introspective Reasoning (ICML 2025 · similarity 0.92)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.92)
- Safety Alignment Can Be Not Superficial With Explicit Safety Signals (ICML 2025 · similarity 0.92)
- Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models (ICLR 2025 · similarity 0.91)
- Safety Layers in Aligned Large Language Models: The Key to LLM Security (ICLR 2025 · similarity 0.91)
- RSafe: Incentivizing proactive reasoning to build robust and adaptive LLM safeguards (NeurIPS 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.
