Research map / Large language models / Retrieval and knowledge
Editing as Unlearning: Are Knowledge Editing Methods Strong Baselines for Large Language Model Unlearning?
Zexi Li, Xiangzhu Wang, William F. Shen, Meghdad Kurmanji, Xinchi Qiu, Dongqi Cai, Chao Wu, Nicholas D. Lane
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 editing methods (WISE, AlphaEdit) instead of unlearning baselines for pretrained knowledge | LLM unlearning · TOFU factual, Llama2-7B-chat | 1 - Rouge1@Forget + Rouge1@Retain: GD 1.321, DPO 1.086 -> WISE 1.481, AlphaEdit 1.606 | Improved |
| use editing methods instead of unlearning baselines for finetuned knowledge | LLM unlearning · PISTOL, Mistral-7B-instruct | 1 - Rouge1@Forget + Rouge1@Retain: AlphaEdit 1.074, WISE 1.309 vs GD 1.544, GA 1.493 | Mixed |
| add self-improvement pipeline to WISE and AlphaEdit (ablation) | WISE, AlphaEdit · TOFU factual, Llama2-7B-chat | Entailment (human rating): AlphaEdit 1.73 -> 2.12 | Improved |
| add self-improvement pipeline to WISE (ablation) | WISE · TOFU factual, Llama2-7B-chat | Trustworthiness (human rating): slight decline (-0.230) | Hurt |
| add self-improvement pipeline to AlphaEdit (ablation) | AlphaEdit · TOFU factual, Llama2-7B-chat | Rephrase (1 - Rouge1): 0.68 -> 0.90 | Improved |
| add query merging to ROME and MEMIT (ablation) | ROME, MEMIT · TOFU factual, 80 forget samples, Llama2-7B-chat | 1 - Rouge1@Forget + Rouge1@Retain: ROME/MEMIT 1.622/1.632 at 5 merged queries, vs DPO 1.123, GD 1.596 | Improved |
| merge more queries into one (increase n) (ablation) | ROME, MEMIT with query merging · TOFU factual, 80 forget samples | 1 - Rouge1@Forget + Rouge1@Retain: peak at 5 merged queries | Mixed |
| increase number of forget samples from 20 to 100 (ablation) | WISE, AlphaEdit · TOFU factual, Llama2-7B-chat | 1 - Rouge1 on forget set | Improved |
| rephrase the forget queries (generalization attack) (ablation) | all methods · TOFU factual, 40 forget samples, Llama2-7B-chat | Forget Rouge1: AlphaEdit 0.08 -> 0.14; WISE 0.28 -> 0.46 | Hurt |
| use ROME or MEMIT directly for unlearning without query merging | ROME, MEMIT · TOFU factual, Llama2-7B-chat | Retain Rouge1: MEMIT retain 0.01, ROME 0.04 | Hurt |
| use GRACE for unlearning | GRACE · TOFU factual, Llama2-7B-chat | Forget Rouge1: 0.65 forget Rouge1 (0.82 before) | No clear effect |
Closest papers by meaning
- UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models (NeurIPS 2025 · similarity 0.93)
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.92)
- Uncovering Overfitting in Large Language Model Editing (ICLR 2025 · similarity 0.92)
- In-Context Unlearning: Language Models as Few-Shot Unlearners (ICML 2024 · similarity 0.91)
- Catastrophic Failure of LLM Unlearning via Quantization (ICLR 2025 · similarity 0.91)
- Unveiling the Pitfalls of Knowledge Editing for Large Language Models (ICLR 2024 · similarity 0.91)
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.
