Training-Free Halving of Activated Experts in Fine-Grained Mixture-of-Experts Models
Modern fine-grained Mixture-of-Experts (MoE) models route each token to a small number of experts and renormalize their router probabilities. We show that this renormalization implicitly calibrates expert output gain to the training top-$k$: reducing $k$ at inference changes not only which experts are used but also the strength of the expert branch. We separate these effects by activating the top $k_1$ experts while normalizing by the probability mass of the top $k_2$ experts, introducing one integer with no parameters, training, or measurable compute overhead. On Qwen3.6-35B-A3B, reducing from 8 to 4 experts causes a 4.65-point MMLU drop under standard renormalization but only 0.35 point...
| Authors | Xing Chen, Hengshuai Yao |
|---|---|
| Primary category | cs.LG |
| Published | 4 September 2026 |
| Last updated at source | 4 September 2026 |
Connected in the knowledge graph
- Davd-b01/qwen3.8-flash-next-40b-prune-researchModel · cited by
Models and datasets referencing this paper
Silytics shows public metadata from arXiv and links to the original. Information may be incomplete or out of date. Last refreshed 17 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.