AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

Tracing and Reversing Edits in LLMs

Paul Youssef, Zhixue Zhao, Christin Seifert, Jörg Schlötterer

ICLR 2026 · Large language models › Retrieval and knowledge · approach: Knowledge editing in LLMs · 4 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
exclude more top singular components when rebuilding the edited matrix (k=0 to best k) (ablation)Bottom-rank approximation, ROME edits · CounterFact edit reversalreversal accuracy: 0.00 -> 94.52 (GPT2-XL, k=11)Improved
use bottom-rank approximation (best k) instead of edited weightsROME-edited models · KL divergence to original modelKL divergence: 11.567 -> 0.218 (GPT-J); 8.988 -> 1.534 (Qwen2.5)Improved
apply reversal with k=15 to edited Llama-3Llama-3-8B ROME-edited · six GLUE tasks incl. MMLUtask performanceNo clear effect
reverse edits with only k=1 on Llama-3 instead of GPT2-XL (ablation)Bottom-rank approximation · CounterFact edit reversalreversal accuracy: 87.10 (GPT2-XL) vs 5.48 (Llama-3) at k=1Mixed
train layers further from the edited layer in EditScope (ablation)EditScope · CounterFact edited-object tracingedited object accuracyHurt
replace ROME with r-ROME editsEditScope · CounterFact edited-object tracingedited object accuracy: differences small (< 2.7 p.p.)Hurt
apply EditScope to AlphaEdit instead of ROME editsEditScope · CounterFact edited-object tracingedited object accuracy: 96.47 -> 2.83 (Llama-3)Hurt
apply EditScope to MEND edits instead of ROMEEditScope · CounterFact edited-object tracingedited object accuracy: 97.60 -> 99.72 (GPT-J)Improved
reverse MEMIT instead of ROME editsBottom-rank approximation, Qwen2.5-7B · CounterFact edit reversalreversal accuracy: 62.90 -> 54.88Hurt
count unique predictions across k on edited vs unedited weightsBottom-rank approximation · wikitext-103 edit detectionavg unique predictions: 1.37 -> 2.46 (GPT-J); 1.36 -> 1.84 (Llama-3)Improved

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