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
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 |
|---|---|---|---|
| add COCA-structured data to LoFiT | LoFiT · Jailbreak avg, LLaMA-3.1-8B | attack success rate: 44.9 -> 10.4 | Improved |
| add COCA-structured data to ReFT | ReFT · Jailbreak avg, LLaMA-3.1-8B | attack success rate: 44.3 -> 22.4 | Improved |
| add COCA-structured data to RR | RR · Jailbreak avg, LLaMA-3.1-8B | attack success rate: 7.8 -> 7.0 | No clear effect |
| add COCA data to LoFiT | LoFiT · Math/code utility avg, LLaMA-3.1-8B | pass@1: 43.6 -> 44.3 | Improved |
| add COCA data to ReFT | ReFT · Math/code utility avg, LLaMA-3.1-8B | pass@1: 43.3 -> 43.5 | No clear effect |
| add COCA training to the model | LLaMA-3.1-8B · Linear probe separability | max ROC-AUC: 0.894 -> 0.972 | Improved |
| add COCA SFT supervision to prompt-only template (ablation) | Prompt-Only · Jailbreak OOD avg | attack success rate: 30.9 -> 16.2 | Improved |
| add LoFiT intervention to COCA-SFT model (ablation) | COCA-SFT only · Jailbreak OOD avg | attack success rate: 16.2 -> 10.4 | Improved |
| add explicit concept reasoning instead of a fixed generic concept (ablation) | COCA · Jailbreak attacks | attack success rate | Improved |
| add concept tags in addition to think tag (ablation) | COCA · Jailbreak attacks | attack success rate | Improved |
| add STAIR introspective stage to COCA (ablation) | COCA · OOD jailbreak | attack success rate: 4.7% reduction | Improved |
| increase safety data from 1K to 10K (ablation) | COCA LoFiT · Jailbreak OOD avg | attack success rate: 27.5 -> 21.4 -> 10.4 | Improved |
Closest papers by meaning
- Safety Alignment Can Be Not Superficial With Explicit Safety Signals (ICML 2025 · similarity 0.87)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.87)
- Semantic Representation Attack against Aligned Large Language Models (NeurIPS 2025 · similarity 0.87)
- Safety Reconstructed: Generative Modeling via Masked Diffusion Builds Strong Safety Guardrails (NeurIPS 2026 · similarity 0.86)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.86)
- Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs (CVPR 2026 · similarity 0.86)
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.
