Research map / Large language models / Language, safety and interpretability
Inference-Time Personalized Safety Control via Paired Difference-in-Means Intervention
Tran Huynh, Ruoxi Jia
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 |
|---|---|---|---|
| use paired contrast mean shift instead of direct prompting | LLaMA-3.1-8B · Personalized safety (avg) | harmfulness score: 3.36 -> 1.73 | Improved |
| use paired contrast mean shift instead of direct prompting | LLaMA-3.1-8B · Personalized safety (avg) | utility score: 8.51 -> 7.95 | Hurt |
| use paired prompts (PCMS) instead of unpaired mean shift (UMS) | LLaMA-3.1-8B · Personalized safety (avg) | utility score: 5.80 -> 7.95 | Improved |
| use unpaired mean shift (UMS) instead of PCMS | LLaMA-3.1-8B · Personalized safety (avg) | harmfulness score: 1.73 -> 1.43 | Improved |
| use PCMS instead of ILCS-local | LLaMA-3.1-8B · Personalized safety (avg) | utility score: 7.57 -> 7.95 | Improved |
| use PCMS instead of ILCS-global | LLaMA-3.1-8B · Personalized safety (avg) | harmfulness score: 2.95 -> 1.73 | Improved |
| use PCMS instead of in-context learning | LLaMA-3.1-8B · Personalized safety (avg) | harmfulness score: 2.86 -> 1.73 | Improved |
| use PCMS instead of retrieval-augmented generation | LLaMA-3.1-8B · Personalized safety (avg) | harmfulness score: 2.82 -> 1.73 | Improved |
| apply PCMS on Mistral-7B instead of direct prompting | Mistral-7B · Personalized safety, Violence | harmfulness score: 3.98 -> 2.12 | Improved |
| apply PCMS on DeepSeek-R1-Distilled-LLaMA3-8B instead of direct prompting | DeepSeek-R1-Distilled-LLaMA3-8B · Personalized safety, Violence | harmfulness score: 2.87 -> 1.62 | Improved |
| add PCMS intervention | LLaMA-3.1-8B · GSM8K | accuracy: 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 ablation | harmfulness and utility | Hurt |
Closest papers by meaning
- Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements (ICLR 2025 · similarity 0.88)
- AlphaAlign: Incentivizing Safety Alignment with Extremely Simplified Reinforcement Learning (ICLR 2026 · similarity 0.87)
- On-Policy Consistency Training Improves LLM Safety with Minimal Capability Degradation (NeurIPS 2026 · similarity 0.87)
- CARE: Decoding-Time Safety Alignment via Rollback and Introspection Intervention (NeurIPS 2025 · similarity 0.87)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.86)
- RSafe: Incentivizing proactive reasoning to build robust and adaptive LLM safeguards (NeurIPS 2025 · 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.
