Research map / Architectures and efficiency / Mixture of experts
MoEEdit: Efficient and Routing-Stable Knowledge Editing for Mixture-of-Experts LLMs
Yupu Gu, Rongzhe Wei, Andy Zhu, Pan Li
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 |
|---|---|---|---|
| add per-expert null-space projection (ablation) | MoEEdit · routing similarity, editing set, layers 11-20 | routing similarity (%): 73.64 -> 86.62 | Improved |
| add per-expert null-space projection (ablation) | MoEEdit · routing KL divergence | KL divergence: 0.0834 -> 0.02 | Improved |
| use MoEEdit instead of UnKE | UnKE · COUNTERFACT efficacy, Qwen3-30B-A3B | efficacy (%): 89.30 -> 99.30 | Improved |
| use MoEEdit instead of strongest baseline | AdaLoRA/UnKE · ZsRE efficacy, Qwen3-30B-A3B | efficacy (%): 44.19 -> 84.47 | Improved |
| use MoEEdit instead of FT | FT · COUNTERFACT generalization, GPT-OSS-20B | generalization (%): 58.40 -> 44.10 | Hurt |
| use MoEEdit instead of FT-L | FT-L · COUNTERFACT specificity, Qwen3-30B-A3B | specificity (%): 71.48 -> 80.97 | Improved |
| use MoEEdit instead of AdaLoRA | AdaLoRA · ZsRE specificity, GPT-OSS-20B | specificity (%): 33.60 -> 32.55 | No clear effect |
| use randomized BCD instead of closed-form solver (ablation) | MoEEdit · solver scalability with experts | runtime | Improved |
| increase number of BCD passes (ablation) | MoEEdit · COUNTERFACT editing metrics | efficacy and generalization | No clear effect |
| use MoEEdit instead of dense-model editors | FT, AdaLoRA · routing similarity, Qwen3-30B-A3B | routing similarity (%): 23.57 -> 86.62 | Improved |
Closest papers by meaning
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.90)
- Oracle-MoE: Locality-preserving Routing in the Oracle Space for Memory-constrained Large Language Model Inference (ICML 2025 · similarity 0.89)
- Ada-K Routing: Boosting the Efficiency of MoE-based LLMs (ICLR 2025 · similarity 0.89)
- UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models (NeurIPS 2025 · similarity 0.88)
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.88)
- OpenMoE: An Early Effort on Open Mixture-of-Experts Language Models (ICML 2024 · 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.
