Research map / Adversarial robustness and security / Attacks on text and graphs
Adversarial Reinforcement Learning for Robust Diffusion Large Language Model Unlearning
Zhiwei Zhang, Yudi Lin, Linlin Wu, Fali Wang, Yi Xin, Xiaomin Li, Minhua Lin, Xianfeng Tang, Qi He, Suhang Wang
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 causal influence regularization (ablation) | adversarial RL unlearning (Dream-v0-Instruct-7B) · TOFU (Clean, L128 S128) | LLM-as-Judge unlearning score (lower better): 0.63 -> 0.55 | Improved |
| add causal influence regularization (ablation) | adversarial RL unlearning (Dream-v0-Instruct-7B) · TOFU (Adversarial Prefix, L128 S128) | LLM-as-Judge unlearning score (lower better): 0.70 -> 0.65 | Improved |
| add causal influence regularization (ablation) | adversarial RL unlearning (Dream-v0-Instruct-7B) · TOFU (Adversarial Full, L128 S128) | LLM-as-Judge unlearning score (lower better): 0.75 -> 0.63 | Improved |
| add regret-based difficulty calibration (ablation) | adversarial RL unlearning (Dream-v0-Instruct-7B) · TOFU (Clean, L128 S128) | LLM-as-Judge unlearning score (lower better): 0.58 -> 0.55 | Improved |
| add regret-based difficulty calibration (ablation) | adversarial RL unlearning (Dream-v0-Instruct-7B) · TOFU (Adversarial Prefix, L128 S128) | LLM-as-Judge unlearning score (lower better): 0.73 -> 0.65 | Improved |
| add regret-based difficulty calibration (ablation) | adversarial RL unlearning (Dream-v0-Instruct-7B) · TOFU (Adversarial Full, L128 S128) | LLM-as-Judge unlearning score (lower better): 0.80 -> 0.63 | Improved |
| train unlearning with adversarial RL instead of gradient ascent | LLaDA-8B-Instruct on TOFU and MUSE · TOFU/MUSE adversarial contexts | LLM-as-Judge and ROUGE | Improved |
| lengthen output from short to 128 tokens | GA-unlearned DLM · TOFU | knowledge recall risk | Hurt |
| add adversarial prefix and suffix contexts at evaluation | GA-unlearned DLM · TOFU | unlearning effectiveness | Hurt |
| increase denoising steps | GA-unlearned DLM · TOFU | unlearning effectiveness | No clear effect |
| remove causal influence regularization (training curves) (ablation) | adversarial RL unlearning · DLM training reward | DLM reward | 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 0.91)
- Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models (NeurIPS 2024 · similarity 0.91)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.89)
- Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and Beyond (ICML 2025 · similarity 0.89)
- On Effects of Steering Latent Representation for Large Language Model Unlearning (AAAI 2025 · similarity 0.89)
- Adversarial Mixup Unlearning (ICLR 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.
