Newest papers

SPARE: Sparsity Allocation via Routing Entropy for Efficient MoEs

Yifan Yang, Hanling Yi, Dongquan Yang, Feng Lin, Rong Xiao, Aijun Li, Xiaotian Yu · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.29662 · cited by 0 works (OpenAlex) · open-access PDF

French

Les architectures Mixture-of-Experts (MoE) étalent efficacement la capacité du modèle, mais leur empreinte de mémoire substantielle constitue une barrière majeure pour le déploiement. L'étalage expert réduit cette empreinte, mais la plupart des méthodes se concentrent sur lequel les experts éliminent et étalent chaque couche de manière égale. Nous proposons SPARE (Sparsity Allocation via Routing Entropy), une méthode sans formation, sélecteur-agnostic qui détermine combien d'experts éliminer de chaque couche en utilisant l'entropie de routage, l'entropie des tokens normalisée Top-K poids de porte sur les données de calibrage. Lorsque le poids de la porte d'un jeton est

Machine translation (Meta M2M100); the English original follows.

Abstract

Mixture-of-Experts (MoE) architectures scale model capacity efficiently, yet their substantial memory footprint poses a major barrier to deployment. Expert pruning reduces this footprint, but most methods focus on which experts to remove and prune every layer equally. We propose SPARE (Sparsity Allocation via Routing Entropy), a training-free, selector-agnostic method that decides how many experts to remove from each layer using routing entropy, the entropy of tokens' normalized Top-K gate weights on calibration data. When a token's gate weight is spread evenly, removing one active expert discards little of it, so SPARE prunes layers with more diffuse routing more aggressively. A perturbation bound under random pruning formalizes this intuition, and controlled interventions with a fixed selector link concentrated routing to larger output changes. With the same selector and budget, SPARE improves average accuracy over uniform allocation on Mixtral, Qwen3, and DeepSeek models, especially at high pruning ratios. At 50% expert pruning, it raises the eight-task average of Qwen3-30B-A3B from 57.2% to 63.8%, and the math, code, and science average of DeepSeek-R1 from 80.1 to 82.8.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Spare part, Computer science, Entropy (arrow of time), Mathematical optimization, Pruning, Routing (electronic design automation)

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange