Research map / Adversarial robustness and security / Attacks on text and graphs
Robust LLM Unlearning via Post Judgment and Multi-round Thinking
Xinrui Chen, Xu Cao, Jianhao Zhang, Pinlong Zhao, Di Gao, Ou Wu
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 in-context prompt cleaning (IPC) (ablation) | PoRT · TOFU 10% noise prefix attack | HFQ: 0.38 -> 0.84 | Improved |
| add selective multi-round thinking (SMT) (ablation) | PoRT · TOFU 10% noise prefix attack | HFQ: 0.25 -> 0.84 | Improved |
| use post-judgment on question-answer pairs instead of pre-judgment on raw input (ablation) | PoRT · TOFU 10% noise prefix attack | HFQ: 0.43 -> 0.84 | Improved |
| use post-judgment on question-answer pairs instead of pre-judgment on raw input (ablation) | PoRT · TOFU 10% noise prefix attack | MU: 0.38 -> 0.67 | Improved |
| add in-context prompt cleaning (IPC) (ablation) | PoRT · TOFU 10% noise prefix attack | MU: 0.65 -> 0.67 | No clear effect |
| apply noise prefix attack to input pre-filtering | ECO · TOFU 5% (Llama-2-7b-chat-hf) | forget probability: 0.0018 -> 0.8055 | Hurt |
| apply noise prefix attack to input pre-filtering | GUARDRAIL · TOFU 5% (Llama-2-7b-chat-hf) | forget probability: 0.2327 -> 0.3827 | Hurt |
| apply composite question attack to input pre-filtering | ECO · TOFU 1% | forget probability: 0.0007 -> 0.5215 | Hurt |
| apply composite question attack | GUARDRAIL · TOFU 10% | ROUGE forget: 0.2471 -> 0.5962 | Hurt |
| apply adversarial attacks to input pre-filtering | ECO · WMDP hazardous accuracy | accuracy (%): 24.9 -> 67.0 | Hurt |
| apply noise prefix attack | ECO · WMDP-Bio (Meta-Llama-3-8B) | accuracy (%): 24.5 -> 66.2 | Hurt |
| apply noise prefix attack | ECO · WMDP-Chem (Qwen) | accuracy (%): 24.7 -> 50.0 | Hurt |
Closest papers by meaning
- Towards Robust Knowledge Unlearning: An Adversarial Framework for Assessing and Improving Unlearning Robustness in Large Language Models (AAAI 2025 · similarity 0.92)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.91)
- Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and Beyond (ICML 2025 · similarity 0.91)
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · similarity 0.90)
- LLM Unlearning via Loss Adjustment with Only Forget Data (ICLR 2025 · similarity 0.90)
- Invariance Makes LLM Unlearning Resilient Even to Unanticipated Downstream Fine-Tuning (ICML 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.
