Challenge: Existing inference optimizations for coarse-grained Mixture-of-Experts models implicitly assume a fixed activation budget, which is poorly understood.
Approach: They propose a training-free policy that adapts token-level activation using router confidence and entropy while remaining within the model’s original budget.
Outcome: The proposed skipping policy can provide substantial throughput gains, but optimal static schedules vary significantly across models and routing mechanisms.

Similar Papers

Alloc-MoE: Budget-Aware Expert Activation Allocation for Efficient Mixture-of-Experts Inference (2026.acl-long)

Copied to clipboard

Challenge: Existing approaches that reduce expert activations lead to severe model performance degradation.
Approach: They propose a framework that optimizes budget allocation coordinately at layer and token levels to minimize model performance degradation.
Outcome: The proposed framework achieves 1.15 prefill and 1.34 decode speedups on DeepSeek-V2-Lite at half of the original budget.
Efficiently Editing Mixture-of-Experts Models with Compressed Experts (2025.findings-emnlp)

Copied to clipboard

Challenge: Mixture-of-Experts models allow for efficient scaling of large language models . fewer experts reduce computational costs, while more experts improve performance .
Approach: They propose to activate only a subset of experts during training and inference . they propose compressed experts that preserve the most important experts .
Outcome: The proposed approach preserves the most important experts while replacing other auxiliary activated experts with compressed experts.
A Closer Look into Mixture-of-Experts in Large Language Models (2025.findings-naacl)

Copied to clipboard

Challenge: Mixture-of-experts (MoE) architectures are gaining increasing attention for their unique properties and remarkable performance.
Approach: They propose a mixture-of-experts architecture that allows for model scaling without sacrificing computational efficiency.
Outcome: The proposed model increases model size without sacrificing computational efficiency . the proposed model is modular and can be used by a broad spectrum of practitioners .
Not All Experts are Equal: Efficient Expert Pruning and Skipping for Mixture-of-Experts Large Language Models (2024.acl-long)

Copied to clipboard

Challenge: Mixture-of-Experts (MoE) LLMs achieve higher performance with fewer active parameters, but are still difficult to deploy due to their immense parameter sizes.
Approach: They propose expert-level sparsification techniques to enhance the deployment efficiency of large language models by introducing plug-and-play expert pruning and skipping techniques.
Outcome: The proposed methods reduce model sizes and increase inference speed while maintaining satisfactory performance across a wide range of tasks.
Bag of Tricks for Sparse Mixture-of-Experts: A Benchmark Across Reasoning, Efficiency, and Safety (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing benchmarks focus on isolated aspects of MoE, with conflicting conclusions . a lack of consensus on optimal design choices is limiting to specific aspects of the model.
Approach: They propose to evaluate two popular MoE backbones across four dimensions of design choices . they find token-level routing and z-loss regularization improve reasoning performance .
Outcome: The proposed framework evaluates two popular MoE backbones on over eight metrics.
Breaking ReLU Barrier: Generalized MoEfication for Dense Pretrained Models (2024.emnlp-main)

Copied to clipboard

Challenge: Existing methods to convert pretrained dense models to MoEs are limited to ReLU-based models with natural sparsity.
Approach: They propose a G-MoEfication approach for arbitrary dense models where activation sparsity assumptions no longer hold.
Outcome: The proposed method reduces the inference cost associated with dense models by sparsely activating experts.
XMoE: Sparse Models with Fine-grained and Adaptive Expert Selection (2024.findings-acl)

Copied to clipboard

Challenge: XMoE leverages small experts and a threshold-based router to selectively engage only essential parameters.
Approach: They propose a novel MoE that leverages small experts to selectively engage only essential parameters.
Outcome: The proposed model can reduce computation load at MoE layers by over 50% without sacrificing performance.
EAC-MoE: Expert-Selection Aware Compressor for Mixture-of-Experts Large Language Models (2025.acl-long)

Copied to clipboard

Challenge: Mixture-of-Experts (MoE) has demonstrated promising potential in scaling LLMs . however, it is hindered by two critical challenges: substantial GPU memory consumption and low activated parameters.
Approach: They propose an Expert-Selection Aware Compressor for Mixture-of-Experts (MoE) that aligns with the characteristics of MoE from the perspectives of quantization and pruning.
Outcome: The proposed approach significantly reduces memory usage and improves inference speed with minimal performance degradation.
Harder Task Needs More Experts: Dynamic Routing in MoE Models (2024.acl-long)

Copied to clipboard

Challenge: Unlike existing MoE approaches that rely on fixed TopK Routing, our dynamic expert selection framework dynamically allocates experts based on the confidence level in expert selection for each input.
Approach: They propose a dynamic expert selection framework that dynamically allocates experts based on the confidence level in expert selection for each input.
Outcome: The proposed method achieves an average improvement of 0.7% with less than 90% activated parameters and outperforms dense models in QA and machine translation tasks.
LightMoE: Task-Aware Expert Availability Management for Memory-Efficient MoE-LLM Inference (2026.findings-acl)

Copied to clipboard

Challenge: Existing solutions for balancing model accuracy with inference latency are limited due to memory constraints.
Approach: They propose a framework for memory-efficient MoE inference that exploits the functional redundancy and temporal locality of expert activation.
Outcome: The proposed framework improves accuracy-efficiency trade-off by 4.3% over pruned models and 2.4% over dynamic swapping methods while maintaining inference latency comparable to pruned model.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations