Expert-Locality-Aware Decode Routing for MoE Serving

Choi & Cho et al. — Sangjin Choi, Sukmin Cho, Yifan Xiong, Ziyue Yang, Youngjin Kwon, Peng Cheng  ·  KAIST, Microsoft Research, Shanghai Xingyunzhili Artificial Intelligence Institute.
ELDR routes decode requests toward workers holding "warm" overlapping experts, using prefill activations as an early signature — a second, exploitable axis beyond ordinary load balancing.