Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

CryptoMoE: Privacy-Preserving and Scalable Mixture of Experts Inference via Balanced Expert Routing

Authors: Yifan Zhou, Tianshi Xu, Jue Hong, Ye Wu, Meng Li

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on Deep Seek Mo E-16.4B, OLMo E-6.9B, and QWen Mo E-14.3B show that Crypto Mo E achieves 2.8 3.5 end-to-end latency reduction and 2.9 4.3 communication reduction over a dense baseline with minimum accuracy loss.
Researcher Affiliation Academia Yifan Zhou Peking University Tianshi Xu Peking University Jue Hong Independent Researcher Ye Wu Independent Researcher Meng Li* Peking University *Corresponding author: EMAIL
Pseudocode Yes Algorithm 1: Secure Dispatch Protocol Πdispatch Algorithm 2: Secure Combine ProtocolΠcombine
Open Source Code Yes Code is available at: https://github.com/PKU-SEC-Lab/Crypto Mo E.
Open Datasets Yes All the models are evaluated on eight famous zero-shot common sense reasoning tasks, including SIQA [13], OBQA [17], Bool Q [11], ARC-easy, ARC-challenge [16], Hella Swag [14], PIQA [12], and Wino Grande [15].
Dataset Splits No All the models are evaluated on eight famous zero-shot common sense reasoning tasks, including SIQA [13], OBQA [17], Bool Q [11], ARC-easy, ARC-challenge [16], Hella Swag [14], PIQA [12], and Wino Grande [15].
Hardware Specification Yes All the experiments are performed on a machine with an Intel Xeon Platinum 8468 CPU (48 cores and 2.1GHz). We consider two network environments: 1) LAN setting with 3Gbps bandwidth and 0.2ms latency; 2) WAN setting with 400Mbps bandwidth and 40ms latency.
Software Dependencies No We implement Crypto Mo E upon the Secret Flow-SPU framework [47], which is a popular framework for privacy-preserving deep learning.
Experiment Setup Yes We benchmark the accuracy, end-to-end amortized latency, and communication cost of different methods in Table 3, using a batch size of 16 and Crypto Mo Et=2.0. ... The token count t assigned to each expert plays a critical role in balancing accuracy and efficiency. ... We consider two network environments: 1) LAN setting with 3Gbps bandwidth and 0.2ms latency; 2) WAN setting with 400Mbps bandwidth and 40ms latency.