Research map / Architectures and efficiency / Mixture of experts
Robustness of Mixtures of Experts to Feature Noise
Dong Sun, Rahul Nittala, Rebekka Burkholz
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 |
|---|---|---|---|
| use MoE probe instead of Lasso | T5-small linear probe · SST-2 (sigma=2.0) | performance drop (%): 10.78 -> 8.60 | Improved |
| use MoE probe instead of Ridge | T5-small linear probe · CoLA (sigma=2.0) | performance drop (%): 12.39 -> 7.67 | Improved |
| use MoE probe instead of Elastic Net | T5-small linear probe · MNLI (sigma=2.0) | performance drop (%): 9.31 -> 7.98 | Improved |
| use MoE probe instead of Elastic Net | T5-small linear probe · AG News (sigma=2.0) | performance drop (%): 3.92 -> 7.75 | Hurt |
| use sparse routed estimator instead of dense estimator at zero overlap | linear classifier · Synthetic block data | classification accuracy: 0.6059 -> 0.7025 | Improved |
| add cross-block feature overlap alpha up to 1.0 (ablation) | sparse estimator · Synthetic block data | classification accuracy: 0.7025 -> 0.6571 | Hurt |
| add cross-block feature overlap alpha up to 1.0 (ablation) | sparse minus dense gap · Synthetic block data | accuracy gap: +0.0966 -> +0.0912 | No clear effect |
| use sparse V-MoE-B/16 instead of dense ViT-L/16 | image classifier · ImageNet-C clean | top-1 accuracy (%): 84.33 -> 84.82 | Improved |
| use sparse V-MoE-B/16 instead of dense ViT-L/16 | image classifier · ImageNet-C Gaussian severity 3 | top-1 accuracy (%): 74.19 -> 77.91 | Improved |
| use sparse V-MoE-B/16 instead of dense ViT-L/16 | image classifier · ImageNet-C Gaussian severity 5 | top-1 accuracy (%): 46.57 -> 59.65 | Improved |
| use MoE instead of dense FFN with matched total FFN parameters | MiniMind LM · pretraining from scratch | training loss | Improved |
| use MoE instead of dense FFN with matched total FFN parameters | MiniMind LM · pretraining from scratch | validation loss vs samples | Improved |
Closest papers by meaning
- On the Expressive Power of Mixture-of-Experts for Structured Complex Tasks (NeurIPS 2025 · similarity 0.91)
- Optimizing Robustness and Accuracy in Mixture of Experts: A Dual-Model Approach (ICML 2025 · similarity 0.90)
- Mixture of Experts Provably Detect and Learn the Latent Cluster Structure in Gradient-Based Learning (ICML 2025 · similarity 0.89)
- Exploiting Activation Sparsity with Dense to Dynamic-k Mixture-of-Experts Conversion (NeurIPS 2024 · similarity 0.89)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.88)
- From Sparse to Soft Mixtures of Experts (ICLR 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.
