Reproduce “CED-MoE: Reversible Service Tiers from Frozen Expert Routers” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Linxiao Li, Weipeng Cao, Jiongjiong Gu, Tingting Song, Miao Liu, Zhixiang Lu, Fanyue Kong, Zhiying Li and 3 more
The authors introduce CED, which uses one calibrated threshold to defer later experts in a mixture-of-experts model’s router shortlist while preserving the original weights of experts that run. On Qwen3-30B-A3B, a C4-calibrated threshold retained 96.7% of full-budget GSM8K accuracy while reducing routed expert calls by 60.8% under the specified setup; fewer calls do not establish lower latency or generalize across workloads.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.