Research map / Large language models / Retrieval and knowledge
The Labyrinth and the Thread: Rethinking Regularizations in Sequential Knowledge Editing for Large Language Models
Zheng Wang, Kaixuan Zhang, Wanfang Chen, Jingwen Zhang, Xiaonan Lu
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 |
|---|---|---|---|
| keep the K0 K0^T P preservation terms instead of dropping them via null-space simplification (ablation) | sequential editing, Memorize-the-Latest task · GLUE SST (LLaMA-3) | accuracy: 0.000 -> 0.846 | Improved |
| keep the K0 K0^T P preservation terms instead of dropping them via null-space simplification (ablation) | sequential editing, Memorize-the-Latest task · MMLU (LLaMA-3) | accuracy: 0.014 -> 0.548 | Improved |
| add error-correction term (Algorithm 1) to RECT post-processing (ablation) | RECT SE · CounterFact | Generalization: 83.60 -> 94.34 | Improved |
| add error-correction term (Algorithm 1) to RECT post-processing (ablation) | RECT SE · CounterFact | Specificity: 84.86 -> 81.56 | Hurt |
| add error-correction term (Algorithm 1) to PRUNE post-processing (ablation) | PRUNE SE · CounterFact | Efficacy/Generalization/Specificity: Eff. 99.82 -> 99.87 | No clear effect |
| enforce OTE alignment of sequential updates (ablation) | PRUNE SE · CounterFact | Efficacy: 56.30 -> 99.87 | Improved |
| enforce OTE alignment of sequential updates (ablation) | RECT SE · CounterFact | Efficacy: 60.35 -> 99.88 | Improved |
| apply post-processing regularization (RECT) to aggregated update in one-time editing | RECT OTE · CounterFact | editing performance | Hurt |
| add PRUNE regularization on top of aligned MEMIT in one-time editing | MEMIT (aligned) · CounterFact | editing performance | No clear effect |
| use OTE-aligned sequential editing with different regularizers | MEMIT/AlphaEdit/PRUNE/RECT (aligned) · GLUE tasks (LLaMA-3) | F1 | No clear effect |
| enforce OTE-SE alignment on PRUNE and RECT (ablation) | PRUNE, RECT · hidden representation shift (LLaMA-3) | distribution shift | Improved |
Closest papers by meaning
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.92)
- Towards Lifelong Model Editing via Simulating Ideal Editor (ICML 2025 · similarity 0.91)
- In-Context Editing: Learning Knowledge from Self-Induced Distributions (ICLR 2025 · similarity 0.91)
- UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models (NeurIPS 2025 · similarity 0.90)
- Rethinking Residual Distribution in Locate-then-Edit Model Editing (NeurIPS 2025 · similarity 0.90)
- Uncovering Overfitting in Large Language Model Editing (ICLR 2025 · similarity 0.90)
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.
