Research map / Large language models / Instruction and fine-tuning
DRAGON: Guard LLM Unlearning in Context via Negative Detection and Reasoning
Yaxuan Wang, Chris Yuhao Liu, Quan Liu, Jinlong Pang, Wei Wei, Yujia Bao, Yang Liu
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 CoT guard instruction (vs none) (ablation) | DRAGON · TOFU-1% (Llama2-7B-Chat) | Deviation Score: 43.9 -> 21.4 | Improved |
| use guard-model CoT instead of template CoT (ablation) | DRAGON · TOFU-1% (Llama2-7B-Chat) | Deviation Score: 33.5 -> 21.4 | Improved |
| use guard-model CoT instead of short template CoT (ablation) | DRAGON · TOFU-10% (Llama2-7B-Chat) | Deviation Score: 40.3 -> 26.5 | Improved |
| dual-layer detector vs RoBERTa classifier (ablation) | DRAGON · WMDP-bio | detection accuracy: 84.2 -> 98.9 | Improved |
| dual-layer detector vs GPT-4o classifier (ablation) | DRAGON · WMDP-chem | detection accuracy: 100.0 -> 98.3 | Hurt |
| DRAGON vs RMU on Yi-34B WMDP-Bio | DRAGON · WMDP-Bio (Yi-34B-Chat) | ProbAcc: 30.6 -> 31.5 | Hurt |
| DRAGON vs best baseline on WMDP-Bio (Llama3.1-8B) | DRAGON · WMDP-Bio (Llama3.1-8B-Instruct) | ProbAcc: 45.1 -> 26.2 | Improved |
| DRAGON vs best baseline at TOFU-10% | DRAGON · TOFU-10% (Llama2-7B-Chat) | Deviation Score: 23.7 -> 26.5 | Hurt |
| DRAGON vs gradient-based baselines at TOFU-5% | DRAGON · TOFU-5% (Llama2-7B-Chat) | Model utility: 0.0 -> 0.6337 | Improved |
| DRAGON vs NPO-RT in continual unlearning | DRAGON · TOFU continual (Llama2-7B-Chat) | DDS: 0.6621 -> 0.2494 | Improved |
| scale base model size and use instruct variants (ablation) | DRAGON · WMDP/TOFU (Qwen2.5) | ROUGE-L gap | Improved |
Closest papers by meaning
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.91)
- OpenUnlearning: Accelerating LLM Unlearning via Unified Benchmarking of Methods and Metrics (NeurIPS 2025 · similarity 0.91)
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.90)
- Unified Parameter-Efficient Unlearning for LLMs (ICLR 2025 · similarity 0.90)
- Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting (NeurIPS 2025 · similarity 0.90)
- In-Context Unlearning: Language Models as Few-Shot Unlearners (ICML 2024 · similarity 0.90)
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.
