AI Paper Atlas

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

Inference-Time Personalized Safety Control via Paired Difference-in-Means Intervention

Tran Huynh, Ruoxi Jia

ICLR 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
use paired contrast mean shift instead of direct promptingLLaMA-3.1-8B · Personalized safety (avg)harmfulness score: 3.36 -> 1.73Improved
use paired contrast mean shift instead of direct promptingLLaMA-3.1-8B · Personalized safety (avg)utility score: 8.51 -> 7.95Hurt
use paired prompts (PCMS) instead of unpaired mean shift (UMS)LLaMA-3.1-8B · Personalized safety (avg)utility score: 5.80 -> 7.95Improved
use unpaired mean shift (UMS) instead of PCMSLLaMA-3.1-8B · Personalized safety (avg)harmfulness score: 1.73 -> 1.43Improved
use PCMS instead of ILCS-localLLaMA-3.1-8B · Personalized safety (avg)utility score: 7.57 -> 7.95Improved
use PCMS instead of ILCS-globalLLaMA-3.1-8B · Personalized safety (avg)harmfulness score: 2.95 -> 1.73Improved
use PCMS instead of in-context learningLLaMA-3.1-8B · Personalized safety (avg)harmfulness score: 2.86 -> 1.73Improved
use PCMS instead of retrieval-augmented generationLLaMA-3.1-8B · Personalized safety (avg)harmfulness score: 2.82 -> 1.73Improved
apply PCMS on Mistral-7B instead of direct promptingMistral-7B · Personalized safety, Violenceharmfulness score: 3.98 -> 2.12Improved
apply PCMS on DeepSeek-R1-Distilled-LLaMA3-8B instead of direct promptingDeepSeek-R1-Distilled-LLaMA3-8B · Personalized safety, Violenceharmfulness score: 2.87 -> 1.62Improved
add PCMS interventionLLaMA-3.1-8B · GSM8Kaccuracy: 80.23% vs 80.12%No clear effect
move the intervention from layers 12-32 to early layers (4 and 8) (ablation)PCMS on LLaMA-3.1-8B · Layer ablationharmfulness and utilityHurt

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