PhD Studentship: A Multi-Agentic AI Framework for Cognitive Network Autonomy

About the Project Project details:

Next-generation networks are rapidly outscaling the capabilities of traditional management paradigms. While early AI/ML models offered a degree of automation, they are fundamentally limited by a "one model for one task" design philosophy. This approach incurs prohibitive engineering costs and yields brittle solutions with poor generalisation to new network conditions, trapping operators in a cycle of designing bespoke, inflexible models. Large Language Models (LLMs) represent a paradigm shift, offering a path to a more sustainable and intelligent approach. Their emergent capabilities in reasoning, planning, and generalisation, acquired through extensive pre-training, provide the foundation for a true cognitive engine capable of managing network complexity holistically. The imperative is to move beyond task-specific AI and harness these foundational models for genuine network autonomy, realised through a collaborative ecosystem of specialised agents.

This research will address the central challenge hindering this vision: the fundamental incompatibility between text-native LLMs and the operational reality of computer networks. Directly applying LLMs is impeded by three core technical barriers: a large input modality gap, as network data consists of diverse, non-textual formats like time-series metrics, graphs, and scalar values; the inefficiency and unreliability of answer generation, where the default token-by-token prediction mechanism is slow and prone to "hallucinating" physically invalid configurations; and the prohibitive adaptation costs of fine-tuning billion-parameter models to acquire domain-specific networking knowledge, especially for reinforcement learning (RL) tasks that traditionally require costly live environment interaction.

To resolve this, this project proposes the design and validation of Cognitive Network LLM, a novel and principled framework for efficiently adapting a single foundation LLM into a collaborative multi-agent system for the networking domain. The framework consists of three key technical innovations. First, a multimodal encoder will be developed to project heterogeneous network data (e.g., time-series, sequences, graphs) into a token-like embedding space that the LLM can natively process. Second, a task-specific networking head will replace the LLM’s default language modelling head. This module will map the LLM's output features directly to a constrained and valid action space, enabling the generation of reliable answers in a single, efficient inference step. Third, an efficient Data-Driven Low-Rank Adaptation scheme will be employed to efficiently specialise this single foundation model into a team of expert agents (e.g., for security, optimisation, and automation). This method instils domain expertise by freezing the LLM's core parameters and fine-tuning only a small set of low-rank matrices for each agent role, drastically reducing GPU memory and training time while preserving the model's pre-trained knowledge. 

The primary outcome of this research will be a validated open-source framework capable of adapting a single general-purpose LLM into a multi-agent system able to solve diverse and complex networking problems through collaboration. This work will establish a new, more sustainable design philosophy for network intelligence, moving from bespoke model engineering to a "one foundational model, many specialised agents" paradigm. By providing a concrete methodology to bridge the gap between general AI and specialised network operations, this project will accelerate the realisation of truly autonomous cognitive networks essential for the 6G era.

Please direct project specific enquiries to: Dr. Haozhe Wang (h.wang3@exeter.ac.uk) Please ensure you read the entry requirements for the potential programme you are applying for. 

Advert information