Research map / Reinforcement learning / RL from feedback
Reinforcement Unlearning via Group Relative Policy Optimization
Efstratios Zaradoukas, Bardh Prenkaj, Gjergji Kasneci
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 GRPO with intrinsic verifiable reward (PURGE) instead of base model | Phi-3-Mini · RWKU Forget Set AA | recall: -12.02% | Improved |
| PURGE vs NPO on forget quality | PURGE · RWKU Forget Set FB | recall: .428 vs .296 | Hurt |
| PURGE forget set construction vs DPO | PURGE · Forget set tokens | tokens per target: x46 fewer than DPO (615,585) | Improved |
| PURGE forget set construction vs GA/NPO | PURGE · Forget set tokens | tokens per target: x22 fewer than GA and NPO (298,067) | Improved |
| PURGE vs base on fluency | PURGE · RWKU FLU | fluency: 133.58 -> 140.90 | Improved |
| PURGE vs base on utility GA | PURGE · RWKU Utility GA | accuracy: -3.62% | Hurt |
| increase Qwen-2.5 model size 0.5B to 3B (ablation) | PURGE · RWKU forget probes | recall | Hurt |
| scale Qwen-2.5 size on utility (ablation) | PURGE · RWKU utility | utility metrics | No clear effect |
| swap NER-constructing LLM (Claude, Deepseek etc.) (ablation) | PURGE · RWKU Stephen King | forget/utility metrics | No clear effect |
Closest papers by meaning
- RULE: Reinforcement UnLEarning Achieves Forget-retain Pareto Optimality (NeurIPS 2025 · similarity 0.91)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.89)
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · similarity 0.89)
- Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and Beyond (ICML 2025 · similarity 0.88)
- LLM Unlearning via Loss Adjustment with Only Forget Data (ICLR 2025 · similarity 0.88)
- Machine Unlearning Doesn't Do What You Think: Lessons for Generative AI Policy and Research (NeurIPS 2025 · similarity 0.88)
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.
