The paper investigates training‑free pruning of Mixture‑of‑Experts (MoE) models, focusing on how to handle the computation when a router selects a pruned expert. It compares replacement operators from three families on several models (OLMoE, DeepSeek‑V2‑Lite, Moonlight, Granite) while keeping the pruning mask fixed, and introduces a new operator called the tilted novice. The study finds that no single operator consistently dominates, with performance varying by model, pruning level, and evaluation metric, and it proposes a five‑item checklist for reporting results.
arXiv:2608. 07890v1 Announce Type: cross Abstract: Mixture-of-Experts (MoE) models decouple total parameters from per-token compute, but deployment still requires storing every expert.
The paper examines how expert pruning—removing low‑importance experts in Mixture‑of‑Experts models—fails when the router is over‑dispersed, a condition caused by aggressive load‑balancing that spreads tokens nearly uniformly across experts. In this regime, traditional importance signals from router probabilities collapse, making perplexity an unreliable predictor of downstream accuracy; for example, the lowest‑perplexity pruning on gpt‑oss‑20B harms mathematical reasoning while the highest‑perplexity pruning preserves it. To address this, the authors introduce Minimax Expert Score Allocation (MESA), a domain‑aware method that iteratively boosts scores for the most affected domain, achieving minimal worst‑case degradation across domains and outperforming baseline pruning strategies on multiple benchmarks while reducing memory usage.
arXiv:2609. 04575v1 Announce Type: cross Abstract: Modern fine-grained Mixture-of-Experts (MoE) models route each token to a small number of experts and renormalize their router probabilities.
ACE introduces a training‑free, calibration‑free framework for adaptive expert skipping in Mixture‑of‑Experts LLMs. It combines a Global Spectral Proxy that estimates global transformation capacity with a Router‑Conditioned Refinement that builds expert‑specific direction prototypes, enabling the model to skip low‑contribution experts while always keeping the top‑1 expert. Offline computation of expert statistics leaves only lightweight table lookups during inference, and experiments on three MoE‑based LLMs show ACE outperforms static and dynamic baselines, especially at high skipping ratios.
MoRA is a framework for pruning Mixture-of-Experts (MoE) models by learning a router bias for each expert and optimizing it with a language‑modeling loss and a routing‑diversity regularizer. The learned biases sharpen routing distributions to identify critical experts and encourage diverse routing preferences. After pruning, MoRA uses an expert approximation mechanism that approximates the outputs of pruned experts with affine transformations of remaining experts, further improving performance.
arXiv:2609.30465v1 Announce Type: cross Abstract: Mixture-of-experts (MoE) models activate few experts per token but store the full expert pool. Expert pruning reduces this storage burden; at a fixed...
The paper introduces Ban&Pick, a post‑training, plug‑and‑play routing strategy for Sparse Mixture‑of‑Experts large language models. It identifies and reinforces a small group of highly influential experts while dynamically pruning redundant ones, leading to accuracy gains across math, code, and reasoning benchmarks. Experiments on DeepSeek and Qwen3 show notable performance improvements and a 1.25× inference speedup without retraining or architectural changes.
The paper investigates how different expert‑deactivation strategies at inference time—removal, freezing, remaining‑expert renormalization, and unnormalized re‑routing—affect token routing and expert weights. Experiments across five instruction‑tuned checkpoints from four model families, plus a base OLMoE checkpoint, show that the choice of strategy interacts with the training‑time weight rule: renormalization performs best when models were trained with raw router probabilities, while unnormalized re‑routing is superior when trained with normalized weights. A weight‑sum model links loss of expert contributions and branch‑scale changes to the training rule, revealing that mismatches can inflate perplexity by up to 28,740× and that additional renormalization penalties grow as fewer experts are activated.
arXiv:2608. 08853v1 Announce Type: new Abstract: Sparse Mixture-of-Experts (MoE) routers commonly use the same scores both to select experts and to weight their already-computed outputs.
The paper introduces Multi-View Latent MoE, a sparse mixture‑of‑experts architecture that uses multiple independent routers to generate richer selection signals and to modulate expert computation. By aggregating sigmoid scores from these routers, the model selects experts that are strongly favored by one or several routers and mixes their relative scores into each expert’s input, thereby reusing routing preferences for both selection and computation. Experiments on a 24‑layer 10B‑A1B model trained on 708 B tokens show a 1.18‑point improvement in nine‑task downstream mean over the baseline, with the Multi‑Router submodule alone providing a 0.6‑point gain and significantly better alignment between router scores and expert utility. whyItMatters":"The approach demonstrates that leveraging multiple routers can both improve expert selection quality and maintain computational efficiency, leading to measurable downstream performance gains on large‑scale language models."
arXiv:2609.25809v1 Announce Type: new Abstract: Fine-grained mixture-of-experts (MoE) architectures have become a mainstream design for open-weight LLMs, with hundreds of experts and increasingly man...
SPARE (Sparsity Allocation via Routing Entropy) is a training‑free, selector‑agnostic method that determines how many experts to prune from each layer of a Mixture‑of‑Experts model by analyzing the routing entropy of tokens on calibration data. By pruning layers with more diffuse routing more aggressively, SPARE improves accuracy over uniform pruning across several large models, achieving notable gains at high pruning ratios. The approach is supported by a perturbation bound under random pruning and controlled experiments that demonstrate its effectiveness on Mixtral, Qwen3, and DeepSeek models.