Research map / Label-efficient and robust learning / Recognition and long-tail
Toward Calibrated Mixture-of-Experts Under Distribution Shift
Gina Wong, Drew Prinster, Suchi Saria, Rama Chellappa, Anqi Liu
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 entropy-balanced adversarial reweighting (Robust MoE) | soft-routed MoE · CIFAR-10H hard subset | ECE: 0.281 -> 0.074 | Improved |
| Add entropy-balanced adversarial reweighting (Robust MoE) | soft-routed MoE · CivilComments hard subset | ECE: 0.108 -> 0.037 | Improved |
| Add entropy-balanced reweighting | soft-routed MoE · CIFAR-10H | accuracy: 0.920 -> 0.904 | Hurt |
| Calibrate each expert post hoc (MoCaE) | Vanilla MoE · CIFAR-10H hard subset | ECE: 0.281 -> 0.262 | Improved |
| Calibrate each expert post hoc (MoCaE) | Vanilla MoE · CIFAR-10H, CivilComments | accuracy | No clear effect |
| Apply robust term only to routing-relevant examples with ERM term kept (Robust Filtered vs Robust MoE) (ablation) | Robust MoE · CIFAR-10H | ECE: 0.090 -> 0.013 | Improved |
| Apply robust term only to routing-relevant examples (Robust Filtered vs Robust MoE) (ablation) | Robust MoE · CIFAR-10H hard subset | ECE: 0.074 -> 0.122 | Hurt |
| Add temperature scaling after training | Robust methods · CivilComments hard subset | ECE: 0.037 -> 0.029 | Improved |
| Combine FGR with Robust MoE (ablation) | Robust MoE · CivilComments hard subset | ECE: 0.037 -> 0.065 | Hurt |
| Combine FGR with Robust MoE (ablation) | Robust MoE · CIFAR-10H hard subset | ECE: 0.074 -> 0.065 | Improved |
| Add FGR gradient rectification | Vanilla MoE · CivilComments hard subset | ECE: 0.108 -> 0.065 | Improved |
| Add FGR gradient rectification | Vanilla MoE · CIFAR-10H hard subset | ECE: 0.281 -> 0.262 | Improved |
Closest papers by meaning
- Variational Routing: A Scalable Bayesian Framework for Calibrated Mixture-of-Experts Transformers (ICML 2026 · similarity 0.88)
- Reassessing How to Compare and Improve the Calibration of Machine Learning Models (ICLR 2025 · similarity 0.87)
- Algorithms with Calibrated Machine Learning Predictions (ICML 2025 · similarity 0.85)
- Unchosen Experts Can Contribute Too: Unleashing MoE Models’ Power by Self-Contrast (NeurIPS 2024 · similarity 0.85)
- Optimizing Robustness and Accuracy in Mixture of Experts: A Dual-Model Approach (ICML 2025 · similarity 0.85)
- Perturb and Correct: Post-Hoc Ensembles using Affine Redundancy (NeurIPS 2026 · 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.
