Research map / Large language models / Retrieval and knowledge
Multiplicative Orthogonal Sequential Editing for Language Models
Hao-Xiang Xu, Jun-Yu Ma, Ziqi Peng, Yuhao Sun, Zhen-Hua Ling, Jia-Chen Gu
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 |
|---|---|---|---|
| replace additive update with multiplicative orthogonal update (MOSE vs AlphaEdit) | LLaMA3-8B · CounterFact (single-sequential, 4000 edits) | Reliability: 0.9018 -> 0.9887 | Improved |
| replace additive update with multiplicative orthogonal update (MOSE vs AlphaEdit) | LLaMA3-8B · CounterFact (single-sequential, 4000 edits) | Locality: 0.7831 -> 0.8972 | Improved |
| replace additive update with multiplicative orthogonal update (MOSE vs AlphaEdit) | Qwen2.5-7B · CounterFact (single-sequential, 4000 edits) | Reliability: 0.9519 -> 0.9981 | Improved |
| MOSE vs AlphaEdit on conceptual knowledge | LLaMA3-8B · ConceptEdit-Inter (single-sequential, 4000 edits) | Reliability: 0.7012 -> 0.7859 | Improved |
| MOSE vs AlphaEdit in batch-sequential editing | LLaMA3-8B · CounterFact (batch-sequential, 10x500 edits) | Generalization: 0.7835 -> 0.9361 | Improved |
| MOSE vs AlphaEdit in batch-sequential editing | Qwen2.5-7B · CounterFact (batch-sequential, 10x500 edits) | Reliability: 0.9427 -> 0.9775 | Improved |
| move from single to batch-sequential editing under AlphaEdit | LLaMA3-8B · CounterFact | Reliability: 0.9018 -> 0.8222 | Hurt |
| add layer selection algorithm over a predefined layer (ablation) | MOSE batch-sequential · LLaMA3-8B CounterFact | editing performance | Improved |
| also edit neighboring layers beyond the single selected layer (ablation) | MOSE batch-sequential · LLaMA3-8B CounterFact | editing performance | Improved |
| use orthogonal multiplicative update instead of additive update (MOSE vs RECT) | sequential editing · CounterFact t-SNE of hidden states | distribution overlap | Improved |
| MOSE editing time vs AlphaEdit | MOSE · CounterFact | time per sample (s): 18.53 -> 18.60 | No clear effect |
| MOSE editing time vs ROME | MOSE · CounterFact | time per sample (s): 12.09 -> 18.60 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AlphaEdit — AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025) · better on 13 of 14 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Perturbation-Restrained Sequential Model Editing (ICLR 2025 · similarity 0.90)
- Should We Really Edit Language Models? On the Evaluation of Edited Language Models (NeurIPS 2024 · similarity 0.89)
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.88)
- Towards Lifelong Model Editing via Simulating Ideal Editor (ICML 2025 · similarity 0.88)
- UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models (NeurIPS 2025 · similarity 0.88)
- Wasserstein Distance Constraint and Parameter Sparsification for Batched and Iterative Knowledge Editing (AAAI 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.
