AI Research · research

Where Prefill Expert Reduction Hurts MoE Models: A Few Prompt Tokens Carry Most of the Short-Answer ...

论文摘要：Running prompt tokens through fewer routed experts during prefill is a simple way to make mixture-of-experts (MoE) serving cheaper, yet it is unclear what this damages. We loc...

首次发现 2026-10-19T22:08:31Z · 最近更新 2026-10-19T22:08:31Z · 置信度 high · 重要性 7.5

为什么重要

该论文来自一手研究源并通过 3qk7 Research Radar 的新鲜度、来源与研究价值筛选；后续代码、模型权重、复现和相关产品事件可继续关联到该记录。

检测到的变化

研究论文进入 3qk7 高价值 Research Radar

一手证据

arXiv cs.CL · 2026-10-19T22:08:31Z