Research map / Large language models / Language, safety and interpretability
Reverse-Engineering Model Editing on Language Models
Zhiyu Sun, Minrui Luo, Yu Wang, Zhili Chen, Tianxing He
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 white-box weight-update analysis instead of gray-box logit divergence | KSTER subject inference on Llama-3-Instruct · CounterFact AlphaEdit N=100 | top-100 subject recall: 0.45 -> 0.99 | Improved |
| use white-box weight-update analysis instead of gray-box logit divergence | KSTER subject inference on Qwen-2.5-Instruct · zsRE MEMIT N=100 | top-100 subject recall: 0.58 -> 0.98 | Improved |
| use white-box weight-update analysis instead of gray-box logit divergence | KSTER subject inference on GPT-J · CounterFact MEMIT N=10 | top-10 subject recall: 0.96 -> 0.98 | No clear effect |
| increase batch size from 10 to 100 | gray-box baseline on Llama-3-Instruct · zsRE AlphaEdit | subject recall: 0.70 -> 0.43 | Hurt |
| increase batch size from 10 to 100 | KSTER white-box on Llama-3-Instruct · CounterFact MEMIT | subject recall: 1.00 -> 0.99 | No clear effect |
| evaluate prompt recovery on zsRE instead of CounterFact | KSTER prompt recovery on Qwen-2.5-Instruct · MEMIT N=100 | Top-1 prompt recall: 0.47 -> 0.08 | Hurt |
| go from prompt recovery to end-to-end object recovery | KSTER on Llama-3-Instruct · zsRE MEMIT N=100 | Top-1 recall: 0.33 -> 0.07 | Hurt |
| increase camouflage scale alpha from 0 to 5 (ablation) | MEMIT subspace camouflage · CounterFact | average rank of true subject: 50.83 -> 394.12 | Improved |
| increase camouflage scale alpha from 0 to 5 (ablation) | MEMIT subspace camouflage · CounterFact | efficacy: 0.95 -> 0.96 | No clear effect |
| increase camouflage scale alpha from 5 to 7 (ablation) | AlphaEdit subspace camouflage · CounterFact | efficacy: 1.00 -> 0.46 | Hurt |
| increase camouflage scale alpha from 5 to 7 (ablation) | MEMIT subspace camouflage · CounterFact | generalization: 0.53 -> 0.42 | Hurt |
| increase camouflage scale alpha from 0 to 3 (ablation) | AlphaEdit subspace camouflage · CounterFact | generalization: 0.51 -> 0.65 | Improved |
Closest papers by meaning
- Stealth edits to large language models (NeurIPS 2024 · similarity 0.88)
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.87)
- BadEdit: Backdooring Large Language Models by Model Editing (ICLR 2024 · similarity 0.87)
- Should We Really Edit Language Models? On the Evaluation of Edited Language Models (NeurIPS 2024 · similarity 0.87)
- Rethinking Residual Distribution in Locate-then-Edit Model Editing (NeurIPS 2025 · similarity 0.87)
- Position: Editing Large Language Models Poses Serious Safety Risks (ICML 2025 · similarity 0.87)
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.
