Research map / Architectures and efficiency / Mixture of experts
Rewiring Experts on the Fly: Continuous Rerouting for Better Online Adaptation in Mixture-of-Expert Models
Guinan Su, Yanwu Yang, Li Shen, Lu Yin, Shiwei Liu, Jonas Geiping
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 test-time router rerouting with confidence-selected layers | DeepSeek-V2-Lite · Five-benchmark average | average accuracy: 50.89 -> 53.59 | Improved |
| add C3PO reference-based rerouting (100 references) | DeepSeek-V2-Lite · Five-benchmark average | average accuracy: 50.89 -> 48.10 | Hurt |
| add 3-shot or 5-shot in-context learning | DeepSeek-V2-Lite · Five-benchmark average | average accuracy: 50.89 -> 49.26 (3-shot), 49.08 (5-shot) | Hurt |
| select layers by router confidence instead of randomly (ablation) | Rewiring · Five-benchmark average | average accuracy: 49.77 -> 53.59 | Improved |
| update only high-confidence layers instead of all layers (ablation) | Rewiring · Five-benchmark average | average accuracy: 50.68 -> 53.59 | Improved |
| add continuous re-optimization every 128 generated tokens (ablation) | Rewiring · Five-benchmark average | average accuracy: 52.55 -> 53.59 | Improved |
| add rewiring to 3-shot ICL | ICL (3-shot) · Five-benchmark average | average accuracy: 49.26 -> 53.05 | Improved |
| add rewiring to self-consistency (3 paths) | Self-Consistency (3) · Five-benchmark average | average accuracy: 54.88 -> 57.09 | Improved |
| add rewiring to GPT-OSS-20B | GPT-OSS-20B · AIME24 | average correctness (%): 69.58 -> 73.75 | Improved |
| add rewiring to OLMoE | OLMoE · MMLU-redux | accuracy: 37.17 -> 38.30 | No clear effect |
| add rewiring to sample generation | DeepSeek-V2-Lite · HumanEval 10 samples | cosine diversity: 0.390 -> 0.380 | No clear effect |
| re-optimize more frequently (every 96 tokens) or never (ablation) | Rewiring · Five benchmarks | accuracy | Hurt |
Closest papers by meaning
- Unchosen Experts Can Contribute Too: Unleashing MoE Models’ Power by Self-Contrast (NeurIPS 2024 · similarity 0.89)
- Tight Clusters Make Specialized Experts (ICLR 2025 · similarity 0.89)
- Layerwise Recurrent Router for Mixture-of-Experts (ICLR 2025 · similarity 0.89)
- Routing Experts: Learning to Route Dynamic Experts in Existing Multi-modal Large Language Models (ICLR 2025 · similarity 0.89)
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.89)
- Autonomy-of-Experts Models (ICML 2025 · similarity 0.89)
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.
