BEAM: Binary Expert Activation Masking for Dynamic Routing in MoE

Juntong Wu, Jialiang Cheng, Qishen Yin, Yue Dai, Yuliang Yan, Fuyu Lv, Ou Dan, Li Yuan

arXiv:2605.14438 · 2026-05-15 공개 · arXiv · PDF

llm inference-efficiency vllm throughput sparsity dynamic-routing moe binary-mask

Abstract

Mixture-of-Experts (MoE) architectures enhance the efficiency of large language models by activating only a subset of experts per token. However, standard MoE employs a fixed Top-K routing strategy, leading to redundant computation and suboptimal inference latency. Existing acceleration methods either require costly retraining with architectural changes or suffer from severe performance drop at high sparsity due to train-inference mismatch. To address these limitations, we propose BEAM (Binary Expert Activation Masking), a novel method that learns token-adaptive expert selection via trainable binary masks. With a straight-through estimator and an auxiliary regularization loss, BEAM induces dynamic expert sparsity through end-to-end training while maintaining model capability. We further implement an efficient custom CUDA kernel for BEAM, ensuring seamless integration with the vLLM inference framework. Experiments show that BEAM retains over 98\% of the original model's performance while reducing MoE layer FLOPs by up to 85\%, achieving up to 2.5times faster decoding and 1.4times higher throughput, demonstrating its effectiveness as a practical, plug-and-play solution for efficient MoE inference.

한국어 요약

📋 한 줄 요약

**[MoE 효율 추론 / 동적 라우팅]** 학습 가능한 이진 마스크로 토큰 적응적 전문가 선택을 학습해 MoE LLM 추론을 가속하는 BEAM 제안.

🎯 핵심 기여도

💡 핵심 아이디어

"토큰마다 필요한 전문가 수가 다르다"는 직관을 학습 가능한 이진 마스크로 명시화하고, STE로 미분 가능하게 학습하면 train-inference 불일치 없이 동적 sparsity를 얻을 수 있다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: 별도의 아키텍처 변경 없이 기존 MoE 모델에 plug-and-play로 적용 가능한 실용적 추론 가속 솔루션 제공. **한계**: 매우 높은 sparsity에서 일부 도메인의 미세 성능 저하 가능성, 학습 추가 비용은 존재.

🚀 실용적 활용