A comprehensive systematic review published in the International Journal of Networked and Distributed Computing has evaluated integrating large language models into dynamic network reconfiguration to foster sustainable telecommunication systems. Conducted by researchers Manjeet Singh, Abhay Bindle, Preeti Singla, Karan Kumar, and Avtar Singh, the study examines research published between 2020 and April 2025 alongside foundational model developments to evaluate how artificial intelligence can optimize resource allocation across telecommunications infrastructure. Digital networks serve as the backbone of global connectivity, driving economic growth and supporting vast mobile and smart device deployments. However, rapid surges in data consumption require modern networks to adapt dynamically to shifting user demands while ensuring uninterrupted service delivery. At the same time, expanding environmental concerns compel operators to seek low-carbon solutions that balance high performance with ecological sustainability.
The expanding scale and operational complexity of telecommunications infrastructure have made environmental sustainability a central concern across the information and communication technology sector. Estimates indicate that the broader sector currently accounts for two to three percent of global carbon dioxide emissions, a proportion projected to rise as network deployments expand. A primary operational challenge stems from inefficient energy expenditure within physical infrastructure, where hardware components such as base stations, switches, routers, and data processing centers consume substantial electricity even during sustained intervals of low traffic. Conventional network management relies heavily on rigid configurations that remain fixed regardless of shifting demand. Furthermore, commercial service agreements requiring continuous network availability, low latency, and fault resilience often prompt operators to keep redundant hardware components and failover systems active around the clock, generating avoidable energy overhead and carbon emissions.
Within wireless telecommunications systems, radio access networks represent the single largest consumer of operational power, accounting for sixty to eighty percent of total electrical consumption in mobile communications. This energy inefficiency is particularly acute across rural and sparsely populated territories, where base stations and antenna equipment remain fully energized continuously despite serving minimal user traffic during off-peak hours. Addressing these challenges necessitates dynamic network reconfiguration, an adaptive paradigm where operational parameters, topologies, or routing pathways change in real time to satisfy performance, reliability, and environmental goals. Although software-defined networking and network function virtualization decouple control logic from data forwarding hardware and enable virtualization on commodity servers, managing these programmable environments manually or through rigid heuristic scripts remains slow and error-prone. Consequently, operators face mounting operational complexity in coordinating large, heterogeneous networks without introducing service degradation or policy conflicts.
To evaluate how artificial intelligence can address these challenges, the authors implemented a structured literature screening process adhering to Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines. The initial retrieval queried scientific repositories, including IEEE Xplore, ScienceDirect, SpringerLink, and Scopus, yielding 716 academic publications. Following the removal of 26 duplicate entries, 690 remaining records underwent title and abstract screening. From this corpus, 202 publications were shortlisted for in-depth full-text assessment against explicit eligibility criteria. Studies qualified for inclusion if they examined large language models or machine learning methods in telecommunications or network reconfiguration, focused on dynamic, autonomous, or sustainable operations, and presented full-text access with verifiable methodologies published between 2020 and April 2025. Purely theoretical articles lacking validation, non-technical commentaries, and book chapters were excluded. Ultimately, 74 peer-reviewed studies satisfied all quality criteria—evaluating methodological rigor, experimental validation, sustainability relevance, and implementation clarity—and were selected for qualitative and quantitative synthesis.
The synthesized literature reveals that prominent large language models, including Generative Pre-trained Transformer architectures, Pathways Language Model, and Meta’s LLaMA series, possess capabilities that bridge human operational intent and automated network execution. When integrated into intent-based networking and software-defined networking frameworks, these models act as intelligent intermediaries. Rather than requiring network administrators to compose intricate scripts manually, language models can interpret natural language declarations of high-level operational goals. By leveraging token embeddings, attention mechanisms, and few-shot learning, the models parse operator instructions alongside real-time network telemetry, usage patterns, and historical logs. The models subsequently generate structured configuration artifacts, such as JSON or YAML files and OpenFlow rules, tailored to software-defined controllers. Before deployment, these configurations undergo semantic verification and policy compliance checks, allowing the system to enact real-time adaptations to traffic fluctuations and energy goals while substantially lowering administrative workloads.
Specific technical frameworks identified in the systematic review illustrate the effectiveness of language models in intent translation and configuration management. For instance, the Hermes architecture interprets natural language intents to produce controller directives, supporting multi-step reconfigurations while ensuring strict alignment with predefined administrative rules. In related experimental evaluations, language models trained via few-shot learning accurately translated operational goals, such as prioritizing video traffic across specific geographic regions, directly into corresponding OpenFlow directives. In addition, transformer architectures comparable to T5 have been deployed to generate configuration templates for software-defined networking and network function virtualization orchestration platforms, demonstrating superior adaptability, coverage, and operator satisfaction over static rule engines in hybrid cloud setups. Mixture-of-experts language model architectures have similarly demonstrated utility by specializing across distinct domains, including security enforcement, configuration synthesis, and quality of service management, facilitating the layered interpretation of compound operational intents.
The systematic review also examined the application of large language models to network fault resilience, specifically within anomaly detection and automated self-healing workflows. Modern communication systems continuously generate multi-format telemetry logs, performance metrics, and device alerts, which can be ingested by language model processing engines. By analyzing unstructured diagnostic records and correlating temporal event signals, language models can identify fault conditions and generate targeted remediation actions, such as traffic rerouting, configuration overrides, or policy resets. In evaluated sixth-generation terrestrial and non-terrestrial network simulations, language models trained on historical incident logs accurately categorized anomaly patterns and suggested remediation actions, reducing mean time to repair by more than 35 percent. Other evaluated systems integrated language models with automated code generators in edge environments, allowing the framework to interpret abnormal metrics, diagnose root causes without human intervention, and produce command-line patches, thereby supporting automated troubleshooting and reducing downtime.
Policy optimization represents an area where integrating language models with reinforcement learning produces significant measurable improvements across network efficiency and sustainability. Reinforcement learning algorithms learn optimal policies through trial and error, adjusting parameters such as bandwidth allocation, routing paths, and power states based on environmental feedback. When paired with language models, numerical reinforcement optimization is enriched by contextual reasoning. In one evaluated hybrid framework, a language model interpreted operator intent while a reinforcement learning agent determined the optimal SDN execution strategy, yielding a 40 percent reduction in configuration errors compared to standalone language models. In energy-focused applications, multi-objective reinforcement learning models for scheduling in network function virtualization environments achieved up to 25 percent greater energy savings than static management policies. Furthermore, reinforcement learning applied to managing sleep schedules in cellular base stations produced energy savings of up to 30 percent while maintaining required quality of service thresholds.
The integration of Retrieval-Augmented Generation architectures provides an effective technical mechanism for knowledge-driven network automation, bridging language intelligence with proprietary telecommunication documentation. Within these frameworks, technical resources—including configuration guides, vendor manuals, internal wikis, and historical incident records—are segmented into chunks and converted into dense vector embeddings stored within vector databases like FAISS, Pinecone, or Milvus. By employing Approximate Nearest Neighbor indexing, product quantization, and top-K filtering, the system rapidly retrieves contextually relevant documentation when querying the language model. In one industrial knowledge management study evaluated in the review, a retrieval-augmented model achieved a mean reciprocal rank of 97.97 percent and a recall of 91.62 percent. In telecommunication operations, domain-adapted assistants built on transformer architectures demonstrated the ability to interpret operational queries and resolve network issues, with one T5-based troubleshooting assistant achieving an 80 percent fault resolution rate without requiring human operational escalation.
Despite the demonstrated potential of language models, the systematic review identifies substantial operational challenges that constrain their deployment in real-time telecommunications environments. A prominent limitation is inference latency; while traffic engineering decisions in software-defined networks often require execution within milliseconds to prevent packet loss, large language models demand considerable computational time and memory resources. Furthermore, the scarcity of domain-specific labeled training data poses generalization challenges, as general language corpora lack the specialized syntax, device logs, and fault semantics characteristic of multi-vendor network ecosystems. Trust and explainability represent another critical hurdle, given that non-deterministic language models can generate hallucinations—syntactically plausible yet technically erroneous configuration commands that could precipitate major service disruptions if executed without validation. Additionally, integrating language models into legacy network environments is complicated by proprietary command sequences, vendor lock-in, and the absence of standardized real-time telemetry interfaces necessary for automated operational feedback loops.
To address these limitations, the review synthesizes several critical research trajectories essential for establishing sustainable, autonomous telecommunication systems. The authors emphasize the necessity of developing lightweight, domain-specific language models fine-tuned on network logs and policy documentation. Utilizing model compression techniques, such as parameter quantization, knowledge distillation, and pruning, can facilitate edge deployment, thereby minimizing inference latency and reducing the computational energy consumption of network controllers. In addition, establishing rigorous policy validation layers—incorporating policy graphs, compliance verification modules, and zero-trust principles—is necessary to guarantee that synthesized network configurations conform to enterprise security policies and operational constraints before execution. Finally, future work points toward carbon-aware orchestration, where artificial intelligence workloads and network processing tasks are dynamically scheduled in alignment with local renewable energy availability, bridging the gap between language model capabilities and global environmental sustainability goals.
Subject of Research: Systematic review of large language models for sustainable dynamic telecommunication network reconfiguration
Article Title: Integrating Large Language Models into Dynamic Network Reconfiguration: A Systematic Review for Sustainable Telecommunication Systems
Article References: Singh, M., Bindle, A., Singla, P., Kumar, K., & Singh, A. (2026). Integrating Large Language Models into Dynamic Network Reconfiguration: A Systematic Review for Sustainable Telecommunication Systems. International Journal of Networked and Distributed Computing, 14(2), Article 28. https://doi.org/10.1007/s44227-026-00106-2
Image Credits: AI Generated
DOI: 10.1007/s44227-026-00106-2
Keywords: telecommunications, large language models, dynamic network reconfiguration, systematic review, software defined networking, network function virtualization, intent based networking, energy efficiency, radio access networks, reinforcement learning, retrieval augmented generation, anomaly detection, self healing networks, PRISMA methodology, green communications, telecom automation, policy optimization
News Source: Denise Maddox. (November 19, 2026). Systematic Review Evaluates Language Models for Sustainable Telecom Network Reconfiguration. Scienmag.



