LLM DomE

Large language models (LLMs), such as GPT-4, can communicate textually on a variety of topics. They are increasingly ubiquitous in private and public spheres, transforming digital workflows. LLMs are used in many domains, applications, and professional fields, in order to simplify the retrieval of facts and access to information, e.g. for creative text writing, to summarise reports, to aid in complex reasoning tasks, and many more. This research project aims to solve two problems with such modern LLMs.

1. While LLMs perform strongly across a broad range of themes and domains, their effectiveness and performance diminishes in more specialised or niche application areas. We aim to develop more robust, transparent, and efficient techniques for adapting LLMs to specific niche domains, while avoiding the complexity of full model fine-tuning and re-training. We will create better domain-centric LLMs that act as tools for more effective, precise, and transparent retrieval of specialist knowledge from large amounts of data or from the LLM directly.

2. Very large LMs, especially proprietary ones, are mostly opaque, which poses more problems for specialised domains that require a higher degree of model understanding and knowledge grounding, such as journalism, where the identification of original sources for text generation with an LLM is crucial; biomedical engineering, where certain decisions can even impact human lives; or HR use cases, where the selection of candidates must be transparent and fair, and comply with specific legal requirements.

To contribute to this project, you should have prior interest or experience in natural language processing, foundation models, knowledge graphs, AI explainability or neuro-symbolic learning.