Research map / Large language models / Retrieval and knowledge
Perturbation-Restrained Sequential Model Editing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PRUNE, a plug-and-play method that restrains large singular values of the cumulative edit update matrix (via log-based clipping) to lower the condition number, thereby preserving LLM general abilities during sequential model editing.
Problem
Sequential model editing with parameter-modifying methods (MEND, ROME, MEMIT) progressively degrades LLM general abilities (reasoning, summarization, QA, NLI) as the number of edits grows; prior work lacks theory and remedies.
Approach
Matrix perturbation theory links the perturbation of stored key-value associations to the condition number of the edited matrix; PRUNE restrains the large singular values of the summed edit update matrix (keeping the original W's top singular value as a cap, with log compression) to reduce the condition number while keeping edit knowledge.
Key results
- With LLaMA-2 (7B) on ConceptEdit, ROME+PRUNE at 100 edits keeps reasoning at 90.78% versus 12.29% for unrestrained ROME.
- PRUNE keeps MEMIT's downstream performance near the unedited model at 100 edits on ZsRE with LLaMA-2 (7B), where unrestrained MEMIT falls to nearly 0.
- Maximum singular value of the summed MEMIT update grows from 23.39 at 100 edits to 935 at 200 edits, supporting the editing-overfitting motivation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply PRUNE multiple times instead of once (ablation) | ROME on GPT-2 XL · CounterFact, 100 edits | Reasoning: 30.12 -> 30.17 | No clear effect |
| Apply PRUNE multiple times instead of once (ablation) | ROME on GPT-2 XL · CounterFact, 100 edits | Summarization: 12.12 -> 11.38 | Hurt |
| Apply PRUNE multiple times instead of once (ablation) | ROME on GPT-2 XL · CounterFact, 100 edits | Open-QA: 22.63 -> 22.99 | No clear effect |
| Apply PRUNE multiple times instead of once (ablation) | ROME on GPT-2 XL · CounterFact, 100 edits | NLI: 49.83 -> 48.83 | Hurt |
| Apply PRUNE multiple times instead of once (ablation) | ROME on GPT-2 XL · CounterFact, 100 edits | Efficacy: 99 -> 100 | Helped |
| Apply PRUNE multiple times instead of once (ablation) | ROME on GPT-2 XL · CounterFact, 100 edits | Generalization: 92 -> 98 | Helped |
| Apply PRUNE multiple times instead of once (ablation) | ROME on GPT-2 XL · CounterFact, 100 edits | Locality: 74.10 -> 76.1 | Helped |
| Apply PRUNE multiple times instead of once (ablation) | ROME on GPT-2 XL · CounterFact, 500 edits | Reasoning: 27.26 -> 32.68 | Helped |
| Apply PRUNE multiple times instead of once (ablation) | ROME on GPT-2 XL · CounterFact, 500 edits | Summarization: 10.78 -> 11.52 | Helped |
| Apply PRUNE multiple times instead of once (ablation) | ROME on GPT-2 XL · CounterFact, 500 edits | Open-QA: 17.86 -> 22.99 | Helped |
| Apply PRUNE multiple times instead of once (ablation) | ROME on GPT-2 XL · CounterFact, 500 edits | NLI: 45.34 -> 51.17 | Helped |
| Apply PRUNE multiple times instead of once (ablation) | ROME on GPT-2 XL · CounterFact, 500 edits | Efficacy: 56.80 -> 84.6 | Helped |
Surprising findings
- Larger models (LLaMA-2 and LLaMA-3) suffered more general-ability degradation than GPT-2 XL under the same number of MEMIT edits.
- PRUNE's preservation of general abilities for MEND only extends to about 20 edits, indicating an upper limit set by the base editing method.
Limitations
- Edits one fact at a time; batch editing not explored.
- Only evaluated on GPT-2 XL, LLaMA-2 7B, LLaMA-3 8B and three editing methods; larger models and more methods left to future work.
- General-ability experiments on ZsRE were only run with LLaMA-2 (7B) due to compute limits.
Experimental setup
Datasets: ZsRE CounterFact ConceptEdit GSM8K SAMSum Natural Questions RTE
Benchmarks: GSM8K SAMSum Natural Questions RTE ZsRE CounterFact ConceptEdit
Metrics: solve rate ROUGE average exact match accuracy efficacy score generalization score locality score instance change
Baselines: MEND ROME MEMIT
Builds on: ROME MEMIT MEND EasyEdit
Closest papers by meaning
- Should We Really Edit Language Models? On the Evaluation of Edited Language Models (NeurIPS 2024 · similarity 0.91)
- Towards Lifelong Model Editing via Simulating Ideal Editor (ICML 2025 · similarity 0.90)
- Neighboring Perturbations of Knowledge Editing on Large Language Models (ICML 2024 · similarity 0.90)
- Unveiling the Pitfalls of Knowledge Editing for Large Language Models (ICLR 2024 · similarity 0.89)
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.89)
- Reinforced Lifelong Editing for Language Models (ICML 2025 · similarity 0.89)
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.