Generative Language Models for Automatic Error Correction and Enhanced Data Network Security
Year: 2026
Type: dissertation
Abstract: Network management represents a significant operational burden for organisations, placing high demands on specialist expertise, time and financial resources. Fast and accurate diagnosis of network faults directly affects service availability and user productivity. This bachelor's thesis presents the design, implementation and experimental evaluation of multi-agent systems using locally deployable generative language models for automatic fault diagnosis in data networks. Three multi-agent system architectures were designed, differing in the number of specialised agents and the approach to dividing the diagnostic task. Each architecture was tested with five language models — GPT-OSS 20B, GPT-OSS 120B, Qwen 2.5, Qwen 3.5 and gemma4. The best result was achieved by the combination of GPT-OSS 120B in the MAS- architecture, which reached a weighted accuracy of 69.0 % and a strict accuracy of 64.3 %. The experiments demonstrated that diagnostic quality depends primarily on the appropriate selection and scope of input data, as an excessive volume of information leads to reduced accuracy. The thesis confirms that modern locally deployable language models represent a viable option for supporting network diagnostics as a first-level analysis layer, without the need to transmit sensitive data outside the organisation's infrastructure.
Author: Lengálová, Viktorie
Language: Czech
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7165501958