Research map / Large language models / Agents, code and math
SOSBENCH: Benchmarking Safety Alignment on Scientific Knowledge
Fengqing Jiang, Fengbo Ma, Zhangchen Xu, Yuetai Li, Zixin Rao, Bhaskar Ramasubramanian, Luyao Niu, Bo Li, Xianyan Chen, Zhen Xiang, Radha Poovendran
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 |
|---|---|---|---|
| evaluate frontier models on science-knowledge prompts instead of generic AdvBench | frontier LLMs · SOSBench | policy violation rate: GPT-4.1 0.503; Deepseek-R1 0.849 | Hurt |
| domain-specific post-training of LLMs | domain-expert models · SOSBench | policy violation rate: BioMistral-7B-SLERP 0.915 | Hurt |
| scale Deepseek-R1-Distill from 1.5B to 70B | R1-Distill · SOSBench | policy violation rate: 0.948 -> 0.878 | Improved |
| scale Gemma-3 parameters | Gemma-3 · SOSBench | policy violation rate | No clear effect |
| scale o4-mini to o3 | OpenAI reasoning models · SOSBench | policy violation rate: 0.312 -> 0.236 | Improved |
| increase visible reasoning budget | Grok-3-mini and Claude-3.7-Sonnet · SOSBench | policy violation rate: Claude-3.7 0.125 -> 0.167; Grok-3-mini 0.738 -> 0.860 | Hurt |
| increase hidden reasoning budget | o4-mini and Gemini-2.5-Flash · SOSBench | policy violation rate: o4-mini 0.323 -> 0.312 | Improved |
| enable thinking mode | Claude-3.7-Sonnet · SOSBench | policy violation rate: 0.427 -> 0.141 | Improved |
| apply RMU unlearning to bio/chem knowledge (ablation) | zephyr-7b-beta · SOSBench | policy violation rate: 0.903 -> 0.896 | No clear effect |
| apply RMU unlearning (ablation) | Mixtral-8x7B-Instruct · SOSBench / MMLU | policy violation rate / MMLU: PVR 0.806 -> 0.775; MMLU 68.2 -> 67.1 | Mixed |
| add GCG transfer jailbreak (ablation) | Llama-4-Maverick · SOSBench-Lite | policy violation rate: 0.28 -> 0.88 | Hurt |
| add multi-turn Crescendo attack (ablation) | GPT-4.1 · SOSBench-Lite | policy violation rate: 0.463 -> 0.927 | Hurt |
Closest papers by meaning
- SafetyPrompts: A Systematic Review of Open Datasets for Evaluating and Improving Large Language Model Safety (AAAI 2025 · similarity 0.90)
- Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements (ICLR 2025 · similarity 0.90)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.90)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.90)
- Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons (NeurIPS 2025 · similarity 0.89)
- CASE-Bench: Context-Aware SafEty Benchmark for Large Language Models (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.
