AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

OSCS: Online Selection with Provable FAR Control for LLM Safety

Zirui Hu, Zheng Zhang, Yingjie Wang, Dacheng Tao

ICML 2026 · Large language models › Language, safety and interpretability · approach: Safety alignment of LLMs

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
replace SCM-md with OSCS-md for FAR controlSCM-md · Yelp BadNets, BERT-baseFAR: 0.090 -> 0.028Improved
replace SCM-badacts with OSCS-badactsSCM-badacts · Yelp BadNets, BERT-baseFAR: 0.116 -> 0.009Improved
replace CUBE with OSCS-mdCUBE · Yelp BadNets, BERT-baseFAR: 0.322 -> 0.028Improved
replace SCM-md with OSCS-md on RoBERTaSCM-md · Yelp BadNets, RoBERTa-baseFAR: 0.209 -> 0.024Improved
replace SCM-md with OSCS-md in jailbreak detectionSCM-md · AdvBench AutoDAN-ga, Mistral-7BFAR: 0.142±0.003 -> 0.044±0.006Improved
replace SCM-ppl with OSCS-ppl in jailbreak detectionSCM-ppl · AdvBench AutoDAN-ga, Llama-3.1-8BFAR: 0.159±0.003 -> 0.011±0.006Improved
replace SCM-md with OSCS-md on Qwen2 jailbreakSCM-md · AdvBench AutoDAN-ga, Qwen2-7BFAR: 0.000±0.000 vs 0.047±0.001No clear effect
replace SCM-ppl with OSCS-ppl: detection power costSCM-ppl · AdvBench AutoDAN-ga, Mistral-7BPower: 0.899 -> 0.711Hurt
replace SCM-badacts with OSCS-badacts: power cost on Yelp BadNetsSCM-badacts · Yelp BadNets, BERT-basePower: 0.995 -> 0.853Hurt
lower FAR threshold alpha to 0.01OSCS · Backdoor detection, BERT-baseFARNo clear effect
run OSCS over longer stream (more observations)OSCS · Yelp RoBERTa-base, over timeFARImproved
swap scoring function from MDS to BadActs in OSCSOSCS · Backdoor detectionFARNo clear effect

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