Research map / Large language models / Instruction and fine-tuning
Efficient Switchable Safety Control in LLMs via Magic-Token-Guided Co-Training
Jianfeng Si, Lin Sun, Zhewen Tan, Xiangzheng Zhang
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 multi-directional (pos/neg/rej) distillation instead of single-direction pos | SFT on Qwen3-8B · Avg(en) | constructive safety score: 77.55 -> 93.03 | Improved |
| use multi-directional distillation instead of single-direction pos | SFT on Qwen3-8B · Avg(zh) | constructive safety score: 66.60 -> 89.13 | Improved |
| add DPO on top of TPos | TPos_en · Avg(en) | constructive safety score: 93.03 -> 97.58 | Improved |
| add DPO on top of TPos | TPos_zh · Avg(zh) | constructive safety score: 89.13 -> 92.80 | Improved |
| use magic-token co-training instead of SFT+DPO | Qwen3-8B · Avg(en) | constructive safety score: 97.58 -> 97.55 | No clear effect |
| use magic-token co-training instead of SFT+DPO | Qwen3-8B · Avg(zh) | constructive safety score: 92.80 -> 95.03 | Improved |
| add valid magic token at inference (vs random token) (ablation) | MTC/MP · Avg(en) | constructive safety score: 90.83 -> 97.45 | Improved |
| add valid magic token at inference (vs no system prompt) (ablation) | MTC/MP · Avg(zh) | constructive safety score: 91.51 -> 95.13 | Improved |
| fuse English and Chinese policies with policy tokens | MTC · Avg(zh) | constructive safety score: 95.03 -> 95.13 | Improved |
| fuse English and Chinese policies with policy tokens | MTC · Avg(en) | constructive safety score: 97.55 -> 97.45 | No clear effect |
| add magic-token co-training of three behaviors | Qwen3-8B · SAM (EA) | Safety Alignment Margin: 0.033 -> 0.131 | Improved |
| add DPO to TPos | TPos_en · SAM (EA) | Safety Alignment Margin: -0.026 -> -0.013 | No clear effect |
Closest papers by meaning
- Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements (ICLR 2025 · similarity 0.91)
- UpSafe℃: Upcycling for Controllable Safety in Large Language Models (NeurIPS 2026 · similarity 0.91)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.90)
- Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback (NeurIPS 2025 · similarity 0.90)
- AlphaAlign: Incentivizing Safety Alignment with Extremely Simplified Reinforcement Learning (ICLR 2026 · similarity 0.90)
- RSafe: Incentivizing proactive reasoning to build robust and adaptive LLM safeguards (NeurIPS 2025 · similarity 0.89)
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.
