AI Paper Atlas

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

AAAI 2026 · Large language models › Instruction and fine-tuning · approach: Safety alignment after fine-tuning · 4 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
use multi-directional (pos/neg/rej) distillation instead of single-direction posSFT on Qwen3-8B · Avg(en)constructive safety score: 77.55 -> 93.03Improved
use multi-directional distillation instead of single-direction posSFT on Qwen3-8B · Avg(zh)constructive safety score: 66.60 -> 89.13Improved
add DPO on top of TPosTPos_en · Avg(en)constructive safety score: 93.03 -> 97.58Improved
add DPO on top of TPosTPos_zh · Avg(zh)constructive safety score: 89.13 -> 92.80Improved
use magic-token co-training instead of SFT+DPOQwen3-8B · Avg(en)constructive safety score: 97.58 -> 97.55No clear effect
use magic-token co-training instead of SFT+DPOQwen3-8B · Avg(zh)constructive safety score: 92.80 -> 95.03Improved
add valid magic token at inference (vs random token) (ablation)MTC/MP · Avg(en)constructive safety score: 90.83 -> 97.45Improved
add valid magic token at inference (vs no system prompt) (ablation)MTC/MP · Avg(zh)constructive safety score: 91.51 -> 95.13Improved
fuse English and Chinese policies with policy tokensMTC · Avg(zh)constructive safety score: 95.03 -> 95.13Improved
fuse English and Chinese policies with policy tokensMTC · Avg(en)constructive safety score: 97.55 -> 97.45No clear effect
add magic-token co-training of three behaviorsQwen3-8B · SAM (EA)Safety Alignment Margin: 0.033 -> 0.131Improved
add DPO to TPosTPos_en · SAM (EA)Safety Alignment Margin: -0.026 -> -0.013No clear effect

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