Research map / Federated learning and privacy / Personalized and heterogeneous FL
MetaMoE: Diversity-Aware Proxy Selection for Privacy-Preserving Mixture-of-Experts Unification
Weisen Jiang, Shuhao Chen, Sinno Jialin Pan
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 relevance-weighted DPP proxy selection (ablation) | MetaMoE · LLaMA-3.1-8B NLP average | accuracy: 80.32 -> 81.59 | Improved |
| add relevance-weighted DPP proxy selection (ablation) | MetaMoE · CLIP ViT-B/16 CV average | accuracy: 95.39 -> 96.24 | Improved |
| add relevance-weighted DPP proxy selection (ablation) | FlexOlmo · LLaMA-3.1-8B NLP average | accuracy: 77.46 -> 79.78 | Improved |
| add relevance-weighted DPP proxy selection (ablation) | FlexOlmo · LLaMA-3.2-3B NLP average | accuracy: 72.50 -> 73.35 | Improved |
| add context-aware router (sequence-level context) (ablation) | MetaMoE · LLaMA-3.2-3B NLP average | accuracy: 72.62 -> 74.42 | Improved |
| add context-aware router (sequence-level context) (ablation) | MetaMoE · CLIP ViT-B/16 CV average | accuracy: 95.94 -> 96.24 | Improved |
| add proxy-aligned expert training (ablation) | MetaMoE · CLIP ViT-B/16 CV average | accuracy: 94.09 -> 96.24 | Improved |
| add proxy-aligned expert training (ablation) | MetaMoE · LLaMA-3.1-8B NLP average | accuracy: 80.99 -> 81.59 | Improved |
| train experts solely on proxy data instead of private plus proxy data (ablation) | MetaMoE · CLIP ViT-B/16 CV | accuracy: substantial accuracy degradation across all domains | Hurt |
| use relevance-weighted DPP proxies with context-aware router and proxy-aligned training vs FlexOlmo | MetaMoE · CLIP ViT-B/32 CV average | accuracy: 92.92 -> 94.52 | Improved |
| swap public dataset Alpaca for OpenOrca | MetaMoE · LLaMA-3.2-3B NLP average | accuracy: 74.42 -> 74.79 | No clear effect |
| MetaMoE vs unrestricted MoE trained on pooled private data | MetaMoE · LLaMA-3.1-8B NLP average | accuracy: 81.34 -> 81.59 | Improved |
Closest papers by meaning
- CryptoMoE: Privacy-Preserving and Scalable Mixture of Experts Inference via Balanced Expert Routing (NeurIPS 2025 · similarity 0.87)
- ReMoE: Fully Differentiable Mixture-of-Experts with ReLU Routing (ICLR 2025 · similarity 0.86)
- Multilinear Mixture of Experts: Scalable Expert Specialization through Factorization (NeurIPS 2024 · similarity 0.85)
- MoDE: A Mixture-of-Experts Model with Mutual Distillation among the Experts (AAAI 2024 · similarity 0.85)
- Domain-Specific Pruning of Large Mixture-of-Experts Models with Few-shot Demonstrations (NeurIPS 2025 · similarity 0.85)
- Unchosen Experts Can Contribute Too: Unleashing MoE Models’ Power by Self-Contrast (NeurIPS 2024 · similarity 0.85)
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.
