AI Paper Atlas

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

ICLR 2026 · Large language models › Agents, code and math · approach: Mathematical and structural reasoning benchmarks

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 changedWhereMeasuredEffect
evaluate frontier models on science-knowledge prompts instead of generic AdvBenchfrontier LLMs · SOSBenchpolicy violation rate: GPT-4.1 0.503; Deepseek-R1 0.849Hurt
domain-specific post-training of LLMsdomain-expert models · SOSBenchpolicy violation rate: BioMistral-7B-SLERP 0.915Hurt
scale Deepseek-R1-Distill from 1.5B to 70BR1-Distill · SOSBenchpolicy violation rate: 0.948 -> 0.878Improved
scale Gemma-3 parametersGemma-3 · SOSBenchpolicy violation rateNo clear effect
scale o4-mini to o3OpenAI reasoning models · SOSBenchpolicy violation rate: 0.312 -> 0.236Improved
increase visible reasoning budgetGrok-3-mini and Claude-3.7-Sonnet · SOSBenchpolicy violation rate: Claude-3.7 0.125 -> 0.167; Grok-3-mini 0.738 -> 0.860Hurt
increase hidden reasoning budgeto4-mini and Gemini-2.5-Flash · SOSBenchpolicy violation rate: o4-mini 0.323 -> 0.312Improved
enable thinking modeClaude-3.7-Sonnet · SOSBenchpolicy violation rate: 0.427 -> 0.141Improved
apply RMU unlearning to bio/chem knowledge (ablation)zephyr-7b-beta · SOSBenchpolicy violation rate: 0.903 -> 0.896No clear effect
apply RMU unlearning (ablation)Mixtral-8x7B-Instruct · SOSBench / MMLUpolicy violation rate / MMLU: PVR 0.806 -> 0.775; MMLU 68.2 -> 67.1Mixed
add GCG transfer jailbreak (ablation)Llama-4-Maverick · SOSBench-Litepolicy violation rate: 0.28 -> 0.88Hurt
add multi-turn Crescendo attack (ablation)GPT-4.1 · SOSBench-Litepolicy violation rate: 0.463 -> 0.927Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy