Research map / Large language models / Retrieval and knowledge
Tracing and Reversing Edits in LLMs
Paul Youssef, Zhixue Zhao, Christin Seifert, Jörg Schlötterer
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 |
|---|---|---|---|
| exclude more top singular components when rebuilding the edited matrix (k=0 to best k) (ablation) | Bottom-rank approximation, ROME edits · CounterFact edit reversal | reversal accuracy: 0.00 -> 94.52 (GPT2-XL, k=11) | Improved |
| use bottom-rank approximation (best k) instead of edited weights | ROME-edited models · KL divergence to original model | KL divergence: 11.567 -> 0.218 (GPT-J); 8.988 -> 1.534 (Qwen2.5) | Improved |
| apply reversal with k=15 to edited Llama-3 | Llama-3-8B ROME-edited · six GLUE tasks incl. MMLU | task performance | No clear effect |
| reverse edits with only k=1 on Llama-3 instead of GPT2-XL (ablation) | Bottom-rank approximation · CounterFact edit reversal | reversal accuracy: 87.10 (GPT2-XL) vs 5.48 (Llama-3) at k=1 | Mixed |
| train layers further from the edited layer in EditScope (ablation) | EditScope · CounterFact edited-object tracing | edited object accuracy | Hurt |
| replace ROME with r-ROME edits | EditScope · CounterFact edited-object tracing | edited object accuracy: differences small (< 2.7 p.p.) | Hurt |
| apply EditScope to AlphaEdit instead of ROME edits | EditScope · CounterFact edited-object tracing | edited object accuracy: 96.47 -> 2.83 (Llama-3) | Hurt |
| apply EditScope to MEND edits instead of ROME | EditScope · CounterFact edited-object tracing | edited object accuracy: 97.60 -> 99.72 (GPT-J) | Improved |
| reverse MEMIT instead of ROME edits | Bottom-rank approximation, Qwen2.5-7B · CounterFact edit reversal | reversal accuracy: 62.90 -> 54.88 | Hurt |
| count unique predictions across k on edited vs unedited weights | Bottom-rank approximation · wikitext-103 edit detection | avg unique predictions: 1.37 -> 2.46 (GPT-J); 1.36 -> 1.84 (Llama-3) | Improved |
Closest papers by meaning
- Exposing the Illusion of Erasure in Knowledge Editing for LLMs (NeurIPS 2026 · similarity 0.91)
- Position: Editing Large Language Models Poses Serious Safety Risks (ICML 2025 · similarity 0.90)
- Stealth edits to large language models (NeurIPS 2024 · similarity 0.89)
- Should We Really Edit Language Models? On the Evaluation of Edited Language Models (NeurIPS 2024 · similarity 0.88)
- UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models (NeurIPS 2025 · similarity 0.88)
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.88)
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.
