Abstract Modern computer networks are large, complex, and prone to misconfigurations which can cause large-scale network incidents. In this report, I will introduce our recent work on network simulation and verification, which address the shortcomings of existing simulators and verifiers in terms of comprehensiveness, scalability, and evolvability. These methods have been successfully applied to real-world scenarios such as large-scale data center networks and wide area networks.

Abstract FiberCop extends Network Digital Twins (NDTs) beyond internal use to federated ecosystems, enabling cross-domain collaboration. Through NDT-as-a-Service, external partners securely run simulations and evaluate performance without affecting live networks. Leveraging AI and semantic technologies, this approach supports intent-driven management, boosts resource efficiency, and accelerates the transition to Autonomous Networks while creating new business opportunities and aligning with standardization.

Abstract Network configuration plays a critical role in ensuring the stability, scalability, and policy compliance of modern computer networks. However, the inherent complexity of configuration tasks has traditionally required the expertise of experienced network engineers. Recent progress in Large Language Models (LLMs) offers promising opportunities to automate and enhance these processes. In this work, we present NetConfBench, a comprehensive and standardized evaluation framework designed to assess the capabilities of LLM agents in realistic, intent-driven network configuration scenarios. NetConfBench leverages mature network emulators to simulate dynamic network environments, supporting end-to-end workflows including perception, decision-making, and action. It features a diverse benchmark of 40 tasks that span various networking technologies, each with ground truth configurations, annotated reasoning traces and expert-cratfed testcases. Through systematic evaluation of six representative LLMs, we uncover critical limitations in current models, particularly in their ability to handle complex reasoning, perform self-verification, and carry out self-correction. Our findings underscore the need for more interpretable and reliable LLM-based solutions and highlight NetConfBench as a foundational platform for future research at the intersection of LLMs and network configuration.

Abstract Network digital twins (NDTs) are virtual representations of physical systems that enable observation, experimentation, and analysis without interfering with the real-world instance. Recent standards and industry efforts (for example work by 3GPP SA5, ETSI ZSM, ITU-T, IETF, TM Forum) have been working on terminology definition, data models and reference architectures, but gaps remain. Then, a synthesis of the recent progress of standards activity related to NDTs is presented, clarifying terminologies and misconceptions, summarizing the use cases, overview the reference architectures, and highlighting gaps and recommended ways forward for standards alignment.

Abstract The digital twin research of AsiaInfo Technologies is first explored, with a comparative analysis of digital twin research in different companies. Industry applications of digital twin are then summarized, especially in telecom software and services, network lifecycle management, low-altitude economy, etc. The proposed Surface Principle for Digital Gemini is then introduced, as well as its applications in different network domains. Finally, the related works of AsiaInfo's Digital Gemini Research Center are briefly introduced.