AI Paper Atlas

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

Concept Concentration for Faithful Representation Intervention

Hongzheng Yang, Yongqiang Chen, Zeyu Qin, Tongliang Liu, Chaowei Xiao, Kun Zhang, Bo Han

ICML 2026 · Large language models › Language, safety and interpretability · approach: Safety alignment of LLMs · 0 citations (OpenAlex)

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
add COCA-structured data to LoFiTLoFiT · Jailbreak avg, LLaMA-3.1-8Battack success rate: 44.9 -> 10.4Improved
add COCA-structured data to ReFTReFT · Jailbreak avg, LLaMA-3.1-8Battack success rate: 44.3 -> 22.4Improved
add COCA-structured data to RRRR · Jailbreak avg, LLaMA-3.1-8Battack success rate: 7.8 -> 7.0No clear effect
add COCA data to LoFiTLoFiT · Math/code utility avg, LLaMA-3.1-8Bpass@1: 43.6 -> 44.3Improved
add COCA data to ReFTReFT · Math/code utility avg, LLaMA-3.1-8Bpass@1: 43.3 -> 43.5No clear effect
add COCA training to the modelLLaMA-3.1-8B · Linear probe separabilitymax ROC-AUC: 0.894 -> 0.972Improved
add COCA SFT supervision to prompt-only template (ablation)Prompt-Only · Jailbreak OOD avgattack success rate: 30.9 -> 16.2Improved
add LoFiT intervention to COCA-SFT model (ablation)COCA-SFT only · Jailbreak OOD avgattack success rate: 16.2 -> 10.4Improved
add explicit concept reasoning instead of a fixed generic concept (ablation)COCA · Jailbreak attacksattack success rateImproved
add concept tags in addition to think tag (ablation)COCA · Jailbreak attacksattack success rateImproved
add STAIR introspective stage to COCA (ablation)COCA · OOD jailbreakattack success rate: 4.7% reductionImproved
increase safety data from 1K to 10K (ablation)COCA LoFiT · Jailbreak OOD avgattack success rate: 27.5 -> 21.4 -> 10.4Improved

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