The STAR project is an initiative to adapt Large Language Models to assist the teaching process.  The specific problem we are addressing is the process of teaching students how to write a quality research proposal based on Grounded Theory principles.

How can an AI assist in the process of teaching this?

We identified two areas where we feel the current state of the art does not provide solutions:

1. Modern large language models are not specifically optimized for the specific attributes of people interacting with them.
2. There is a possibility of made up, or ‘hallucinated’ information if the sources fed into the AI training are not controlled

We therefore set out to use available techniques in order to solve these problems.  We have now administered a VARK questionnaire to a cohort of 25 students, and have identified a number of students with different learning styles through the automated questionnaire.  We are now planning to re-administer this questionnaire to a new cohort of students at the start of the new school year.

We have also implemented a version of Llama 2 (which is an open source, pre-trained chatbot) which does multimodal retrieval augmented generation.   This means that we can feed information from known correct sources into the chatbot, and it will attempt to generate content based only on the sources we provide.  We can also set it to prefer the format that best corresponds with what students prefer as their learning style.

We are now working on the final stages of the software that this project has been set up to build, which is an AI assistant that can help students learn how to write a quality research proposal based on known correct information, taking student learning styles into account.