国際会議ICDCN2026で16件の研究発表・受賞
2026年1月に奈良で開催されたICDCN 2026にて計16件の研究発表・2件の受賞
研究発表 受賞
国際会議 · 2026
High-Reliability Network-Distributed MoE-LLM System
Abstract
For the practical deployment of distributed Mixture-of-Experts large language models(MoE-LLM), ensuring resilience is as critical as achieving computational efficiency. This research tackles this pressing challenge by introducing a resilient distributed LLM architecture that dynamically maintains system reliability in large-scale networked environments. The proposed framework develops cooperative routing mechanisms that adaptively account for the reliability of individual Experts, as well as resistance mechanisms that mitigate cascade failures and suppress malicious Expert behaviors. Furthermore, it establishes a quantitative evaluation framework through the introduction of the Resilience Score, a comprehensive metric designed to assess system robustness under fault and attack conditions. Through these contributions, this study aims to provide foundational design principles for constructing trustworthy, fault-tolerant, and high-performance distributed LLM systems, paving the way for their safe and reliable integration into future cyber-physical infrastructures.
