EVENT DOSSIER
On September 7, 2026, arXiv cs.AI published the paper “ACE: Adaptive Calibration-Free Expert Skipping for MoE-based LLMs”. This study introduced a new method called ACE, aimed at optimizing the performance of large language models based on the mixed expert (MoE) architecture. This method uses an adaptive calibration-free expert skipping mechanism to dynamically adjust the inference path without relying on traditional calibration data, thereby significantly reducing computational overhead and latency while maintaining model accuracy.
arXiv:2609.05228v1 提出 ACE，这是一种无需训练、无需校准且保留检查点的框架，用于在 MoE 架构大语言模型中实现 Token 自适应专家跳过。ACE 包含两个组件：Global Spectral Proxy (GSP) 和 Router-Conditioned Refinement (RCR)。实验表明，ACE 在三款 MoE 大语言模型和八个基准测试中优于现有静态及动态基线；在 Qwen3.6-35B-A3B 上以 50% 跳过率时，WikiText-2 perplexity 降低 7.96%，下游平均准确率提升 4.15 个百分点。