Research and Applications of Large Language Models in ICT Operations

The talk titled “Large Language Models in ICT Operations” is organized into four main parts: basic viewpoints on LLMs, the development of natural language processing in ICT, concrete applications of LLMs to ICT operations, and future outlook.

1. Basic Viewpoints on LLMs – It outlines three key perspectives: (1) ChatGPT sparked rapid AI industry growth by demonstrating the commercial potential of large‑scale language models; (2) a new paradigm has emerged that is unstoppable; (3) aligning models with human language is essential for better performance, citing RLHF, safety, and instruction tuning.

2. Development of Natural Language in ICT – Traces the idea of translation as encryption/decryption back to Warren Weaver (1894) and highlights the Transformer’s role in decoding natural language. It describes ICT operation as data‑, algorithm‑, and scenario‑driven intelligent maintenance, focusing on KPI, log, and alarm data. It argues that logs, being natural‑language text, can be better understood with LLMs, enabling a unified framework for multiple tasks.

3. Applications of LLMs in ICT – Introduces the LogAIBox platform, which evolved from traditional task‑based intelligent operations to instruction‑based, LLM‑driven operation intelligence. It reviews several generations of technologies: LogAnomaly, LogParse, LogStamp (task‑based); BigLog and DA‑Parser (pre‑training based). BigLog, pretrained on 78 GB of logs, achieves state‑of‑the‑art results on log parsing, anomaly detection, and fault prediction across multiple public datasets, and shows strong generalization and few‑shot capabilities. DA‑Parser combines pre‑training with domain‑specific parsing without manual labeling.

4. Future Outlook – Envisions large models progressing toward Artificial General Intelligence, providing pure natural‑language interaction, passive self‑reporting of issues, and autonomous agent‑based self‑maintenance. It also stresses the need for controllability, safety checks, and compliance to avoid ethical and legal risks.