AI Paper Atlas

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

ICML 2026 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
use white-box weight-update analysis instead of gray-box logit divergenceKSTER subject inference on Llama-3-Instruct · CounterFact AlphaEdit N=100top-100 subject recall: 0.45 -> 0.99Improved
use white-box weight-update analysis instead of gray-box logit divergenceKSTER subject inference on Qwen-2.5-Instruct · zsRE MEMIT N=100top-100 subject recall: 0.58 -> 0.98Improved
use white-box weight-update analysis instead of gray-box logit divergenceKSTER subject inference on GPT-J · CounterFact MEMIT N=10top-10 subject recall: 0.96 -> 0.98No clear effect
increase batch size from 10 to 100gray-box baseline on Llama-3-Instruct · zsRE AlphaEditsubject recall: 0.70 -> 0.43Hurt
increase batch size from 10 to 100KSTER white-box on Llama-3-Instruct · CounterFact MEMITsubject recall: 1.00 -> 0.99No clear effect
evaluate prompt recovery on zsRE instead of CounterFactKSTER prompt recovery on Qwen-2.5-Instruct · MEMIT N=100Top-1 prompt recall: 0.47 -> 0.08Hurt
go from prompt recovery to end-to-end object recoveryKSTER on Llama-3-Instruct · zsRE MEMIT N=100Top-1 recall: 0.33 -> 0.07Hurt
increase camouflage scale alpha from 0 to 5 (ablation)MEMIT subspace camouflage · CounterFactaverage rank of true subject: 50.83 -> 394.12Improved
increase camouflage scale alpha from 0 to 5 (ablation)MEMIT subspace camouflage · CounterFactefficacy: 0.95 -> 0.96No clear effect
increase camouflage scale alpha from 5 to 7 (ablation)AlphaEdit subspace camouflage · CounterFactefficacy: 1.00 -> 0.46Hurt
increase camouflage scale alpha from 5 to 7 (ablation)MEMIT subspace camouflage · CounterFactgeneralization: 0.53 -> 0.42Hurt
increase camouflage scale alpha from 0 to 3 (ablation)AlphaEdit subspace camouflage · CounterFactgeneralization: 0.51 -> 0.65Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy