Approximately 5 months remain of my PhD studies before my dissertation in September. Before this, I have two main research projects that I wish to complete. As both of these projects involve larger collaborations for which the collaborators also depend on my cluster allocation, I find that my current resource allocation is insufficient (for previous research projects I have not needed more hours as I have mainly been using the cluster allocation for my own experiments, without external collaborators sharing on the same resource). Therefore, I submit this project proposal to ask for an extension in GPU-hours for the remaining months of my PhD studies, as I otherwise fear that I will be unable to complete my research projects in time. The two projects for which the GPU-hours would be used are as follows: 1. Knowledge Graph-Augmented Language Models for Improved Consistency and Accuracy in Question Answering This is a thesis project performed by two master's thesis students under supervision by me and an external collaborator at KTH. Since NAISS does not allow compute projects dedicated to only thesis students, we are sharing my cluster project allocation. This project lies in line with my research and we aim to write a paper based on it. In this project, we investigate the potential of using knowledge graphs together with language models (LMs) to improve model accuracy and consistency for question-answering tasks. Previous research projects have mainly focused on leveraging information from unstructured text, such as Wikipedia pages, to improve LM consistency and accuracy. This project aims to investigate whether information from structured knowledge graphs can be leveraged for similar benefits. In addition, we wish to investigate for what situations structured knowledge sources may be preferred over unstructured knowledge sources. This project relies heavily on language models to generate knowledge graph queries from a textual input to fetch relevant information to aid the question answering. 2. A Benchmark of Context Usage Manipulation Techniques This research project is led by me and performed together with researchers based in Copenhagen and Korea. Our intention is to share on my Alvis project for compute and storage resources to simplify the collaborative work. Retrieval-augmented generation (RAG) is a popular approach that helps address the limitations of the memory of language models (LMs). RAG essentially involves having the model consult an external source of knowledge, such as Wikipedia or the Internet, for its output. However, for RAG to work well, it is important that the LM correctly leverages the external information, also referred to as context. Recent research has found that LMs cannot always be trusted to follow provided contexts and many methods for improving LM context utilisation have been proposed. This project aims to benchmark different techniques for improving LM context utilisation, such as prompt engineering, fine-tuning and mechanistic intervention methods across different LMs and datasets. I expect both projects to end around June, so the extension in allocation only needs to last until June, if necessary.