Research map / Architectures and efficiency / Mixture of experts
Opportunistic Expert Activation: Batch-Aware Expert Routing for Faster Decode Without Retraining
Costin-Andrei Oncescu, Qingyang Wu, Wai Tong Chung, Tsai-chuan Wu, Bryan Dev Gopal, Junxiong Wang, Tri Dao, Ben Athiwaratkun
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 piggybacking on top of pruned top-3 routing (ablation) | OEA on Qwen3-30B-A3B · AIME24 | accuracy: 51.2 -> 80.0 | Improved |
| add piggybacking on top of pruned top-3 routing (ablation) | OEA on Qwen3-30B-A3B · LiveCodeBench | accuracy: 37.4 -> 61.2 | Improved |
| add piggybacking on top of pruned top-3 routing (ablation) | OEA on Qwen3-30B-A3B · GPQA | accuracy: 45.7 -> 58.6 | Improved |
| add piggybacking on top of pruned top-5 routing (ablation) | OEA on Qwen3-235B-A22B · LiveCodeBench | accuracy: 53.5 -> 66.1 | Improved |
| add piggybacking on top of pruned top-5 routing (ablation) | OEA on Qwen3-235B-A22B · GPQA | accuracy: 60.6 -> 67.5 | Improved |
| add piggybacking on top of pruned top-3 routing (ablation) | OEA on Qwen3-235B-A22B · AIME24 | accuracy: 17.5 -> 81.4 | Improved |
| add piggybacking at k0=7 (ablation) | OEA on Qwen3-30B-A3B · AIME24 | accuracy: 82.5 -> 78.5 | Hurt |
| reduce experts per token from 8 to 5 without piggybacking | Qwen3-30B-A3B · GPQA | accuracy: 60.2 -> 56.2 | Hurt |
| reduce experts per token from 8 to 5 without piggybacking | Qwen3-235B-A22B · LiveCodeBench | accuracy: 68.5 -> 53.5 | Hurt |
| OEA k0=5 instead of top-8 | Qwen3-235B-A22B · LiveCodeBench | accuracy: 68.5 -> 66.1 | Hurt |
| use OEA k0=3 instead of top-8 | Qwen3-30B-A3B · MoE layer latency | latency (us): 175.7 -> 106.8 | Improved |
| use OEA k0=5 instead of top-8 | Qwen3-30B-A3B · MoE layer latency | latency (us): 175.7 -> 136.0 | Improved |
Closest papers by meaning
- $\textit{Read-ME}$: Refactorizing LLMs as Router-Decoupled Mixture of Experts with System Co-Design (NeurIPS 2024 · 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)
- Layerwise Recurrent Router for Mixture-of-Experts (ICLR 2025 · similarity 0.88)
- MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts (ICLR 2025 · similarity 0.88)
- Certain Head, Uncertain Tail: Expert-Sample for Test-Time Scaling in Fine-Grained MoE (ICML 2026 · 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.
