Research map / Large language models / Language, safety and interpretability
De-attribute to Forget for LLM Unlearning
Xinyang Lu, Jiabao Pan, Rachael Hwee Ling Sim, See-Kiong Ng, Anthony Kum Hoe Tung, Bryan Kian Hsiang Low
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 data de-attribution PPO reward (DareU) instead of gradient ascent | LLM unlearning · TOFU, Llama2-7B-chat | ToW: 0.009 -> 0.816 | Improved |
| use DareU instead of GDiff | LLM unlearning · TOFU, Llama2-7B-chat | ToW: 0.614 -> 0.816 | Improved |
| use DareU instead of NPO | LLM unlearning · TOFU, Qwen3-8B | ToW: 0.638 -> 0.797 | Improved |
| use DareU instead of GDiff | LLM unlearning · ArXiv, Qwen3-8B | ToW*: 0.558 -> 0.660 | Improved |
| use DareU instead of GA (retain-set ROUGE) | LLM unlearning · TOFU, Llama2-7B-chat | DR ROUGE: 0.001 -> 0.797 | Improved |
| use DareU instead of SCRN on truth ratio | LLM unlearning · TOFU, Llama2-7B-chat | Truth Ratio: 0.774 -> 0.564 | Hurt |
| replace LM classifier with sentence-transformer classifier as attribution function (ablation) | DareU · TOFU, Llama2-7B-chat | ToW: 0.816 -> 0.800 | No clear effect |
| replace LM classifier with WASA watermark attribution (ablation) | DareU · TOFU, Llama2-7B-chat | ToW: 0.816 -> 0.806 | No clear effect |
| apply DareU with LM classifier, evaluated by hold-out WASA metric | DareU · TOFU, Llama2-7B-chat | WASA attribution metric (lower is better): 0.854 -> 0.703 | Improved |
| use DareU rather than faster baselines (compute cost) | DareU · ArXiv, Llama2-7B-chat | unlearning time (s): GA 29.83 -> DareU 1595.3 | Hurt |
| use DareU instead of retraining (compute cost) | DareU · ArXiv, Llama2-7B-chat | unlearning time (s): 7852.5 -> 1595.3 | Improved |
| ablate hyperparameters (learning rate, kappa, lambda_dis) (ablation) | DareU · TOFU | unlearning performance | No clear effect |
Closest papers by meaning
- LLM Unlearning via Loss Adjustment with Only Forget Data (ICLR 2025 · similarity 0.93)
- Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning (NeurIPS 2025 · similarity 0.91)
- Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference (NeurIPS 2024 · similarity 0.91)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.90)
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.90)
- RULE: Reinforcement UnLEarning Achieves Forget-retain Pareto Optimality (NeurIPS 2025 · 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.
