The automation of optical networks using large language models (LLMs) is transforming how operators manage complex infrastructures. However, not all LLMs offer the same balance between output quality and computational cost. To address this challenge, HuGLEN emerges as a scalable evaluation pipeline that combines an LLM-based automatic judge with a small set of expert ratings. This approach allows reproducible comparison of different models and assigns them a 'quality efficiency score' (QES) reflecting the trade-off between explanatory quality and inference efficiency.
In the context of optical networks, explainable AI (XAI) systems generate predictions about quality of transmission (QoT). However, these technical outputs need to be translated into operator-friendly explanations, as operators are not always machine learning experts. This is where LLMs come into play: they can generate natural language descriptions, but their performance varies greatly depending on the model family and size. HuGLEN provides a methodology to identify which specific LLM offers the best balance between clarity, accuracy, and response speed.
Experiments conducted with HuGLEN show that a medium-sized model (12 billion parameters) achieves the highest QES, outperforming both small models (with poor explanations) and giant models (with excessive costs). This finding is relevant for companies looking to deploy software process automation in telecommunications networks, where minimizing latency and resources without sacrificing user experience is critical.
From a technical perspective, HuGLEN reduces the human labeling burden by employing an LLM judge system that evaluates generated explanations. Experts only intervene in an initial stage to calibrate quality criteria. This not only accelerates model selection but also ensures consistency in environments where business requirements change frequently. A similar approach can be applied to other domains, such as cloud infrastructure management or cybersecurity, where interpretability of automated decisions is key.
Companies like Q2BSTUDIO, specialized in Artificial Intelligence and custom software development, already work on solutions that integrate LLMs with network monitoring systems. For instance, it is possible to design AI agents that generate automatic reports on network alarms, or conversational assistants that help technicians diagnose incidents. All of this is supported by AWS or Azure cloud for scaling on demand, and with cybersecurity layers that protect sensitive operator data.
LLM evaluation as proposed by HuGLEN fits perfectly into an ecosystem where efficiency is as important as accuracy. Business Intelligence tools (e.g., Power BI) can complement these processes by visualizing model performance in real time, enabling managers to make informed decisions about which version to deploy at each network node. The combination of automation, AI, and BI opens the door to self-managed optical networks, where LLMs act as a natural interface between machines and humans.
In conclusion, HuGLEN represents a practical advance for LLM selection in optical network automation tasks. By prioritizing the balance between quality and cost, companies can adopt language models without incurring unnecessary expenses. Q2BSTUDIO, with its expertise in custom applications, cloud computing, cybersecurity, and AI agents, is ready to help organizations implement such evaluation pipelines and deploy intelligent automation solutions that truly deliver operational value.
