Research Fellow — Multimodal Graph Retrieval Augmented Generation for Creative Assets

The project

Retrieval Augmented Generation (RAG) has emerged as one of the most effective approaches for improving the reliability and factual grounding of Large Language Models (LLMs).

The aim of the project "Multimodal Graph Retrieval Augmented Generation for Creative Assets" is to develop and evaluate a Multimodal Graph Retrieval Augmented Generation (mmGraphRAG) framework capable of performing structured retrieval and reasoning across heterogeneous data modalities.

The project will address the following research questions:

• Multimodal Knowledge Representation
• Graph and Ontological Modelling
• Integration of Vision-Language Models
• Unified Multimodal Embeddings
• Evaluation of Multimodal GraphRAG Systems

The project involves collaboration with Riot Games, one of the world's leading video game companies and the developer of globally successful titles such as League of Legends and Valorant.

The role

This is a research position with a substantial system building component. You will formulate hypotheses, design the experiments that test them, interpret results critically, and publish. 

As a Research Fellow in this project, you will:

Investigate multimodal knowledge representation, and build the pipeline that realizes it. Extract information from text documents, images, tables and figures, and video, and load it into a single structured knowledge base. Open questions may be: how the system decides that references in different modalities denote the same entity, which relationship types are expressive enough without becoming unusable, and how confidence and source information should be represented so that later reasoning can act on them. 

Develop and evaluate retrieval and reasoning methods. Combine similarity-based search over content with navigation through the knowledge base, extending it to chain several reasoning steps. A core contribution is making the system reason explicitly about its uncertainty, choosing between exact and approximate answers according to the evidence. 

Study grounded generation with traceable provenance. Build the mechanisms that keep generated answers consistent with the entities and relationships in the knowledge base, handle contradictory evidence, and attach source references. Design ablation studies isolating each component's contribution to accuracy.

Establish evaluation methodology. Build reproducible evaluation pipelines and the infrastructure that supports them, and package demonstration versions for trial evaluation with the industry partner.

Publish, and supervise. Contribute to papers and present the work. Supervise undergraduate students on self-contained sub-projects, review their code, and integrate their output. Maintain technical documentation and onboard new team members.

Expected profile

You hold a PhD (or expect to complete one shortly) in Computer Science or a closely related discipline. Relevant backgrounds include machine learning, natural language processing, computer vision, information retrieval, knowledge representation and reasoning, data management, or multimodal learning.