PhD position in Large Language Models for education

We have an open Phd position that is part of a large interdisciplinary research project. Collaborating research groups include the Leuven Engineering and Science Education Center (supervisor Tinne De Laet) and the Linguistics Research Unit (supervisor Tim Van de Cruys).
Website unit

Project

The objective of the interdisciplinary project is to advance the field of generative Artificial Intelligence (genAI) and explainable AI (XAI) within the Learning Analytics (LA) domain. In particular, the goal is to facilitate the adoption of AI in higher education by generating and designing explanations of AI outputs for students and teachers using a combination of open-source Large Language Models (LLMs), (interactive) visual analytics and machine learning (ML). As the future of AI should be perceived from a “hybrid intelligence” (Molenaar, 2022) or social perspective in which AI is seen as a participant in conversations for learning (Sharples, 2023), the project aims to gain better insight into these conversations to improve the quality of learning and education.

The project will specifically focus on strengthening the AI algorithms, and the LLMs in particular, in order to support students and teachers as collaborators with genAI and XAI. Also in approaches to strengthen LLMs, explanations play a key role as research has shown hybrid explanations, either originating from humans or AI algorithms, have the potential to improve (open source) LLMs (Li et al., 2022). In the project explanations therefore play a double role:

• Explanations generated by AI algorithms and humans, so-called “hybrid explanations”, fed to an open-source LLM to improve the performance of the LLM.
• Explanations generated by AI algorithms, offered to students and teachers, regarding outputs of AI algorithms to foster adoption of AI towards “hybrid intelligence” in education and improve quality of education.

The focus of the PhD position will be the following research question: How can hybrid explanations provided by humans (domain/educational experts including teachers) and AI algorithms improve the performance of open source genAI algorithms and LLMs in particular.

Currently the research on open-source LLMs is high on the agenda aiming at providing an alternative to (commercially or closed-source) available LLMs, where questions have been raised concerning their transparency, ethical values underlying the models, sustainability, dependency on big-tech, sustainability, climate impact, etc. Open-source LLMs have the potential to be custom-trained in a transparent and targeted manner, giving back control to the developers and owners of the LLMs. The conceptual innovation that this project brings is the improvement of the performance and behavior of open-source LLMs by using hybrid explanations, i.e. both AI-generated and human-generated (domain-expert) explanations. First, the use of the domain-expert explanations to feed LLMs with domain-specific insights and theories is conceptually innovative as it will allow to connect genAI, which are currently considered black-box, to so-called theory-driven LA, which builds on educational theories and models (Dawson et al., 2015). Moreover, it will offer the customization of the LLMs that higher education institutions are looking for such that they can use generative AI in a more transparent and ethical way. Second, the use of AI explanations to improve the LLM itself is innovative. This will open not only a route where AI models are improving each other (peer-improvement) (Li et al., 2022), but also themselves (self-improvement) (Fernando et al., 2023).

Dawson, S., Mirriahi, N., & Gasevic, D. (2015). Importance of Theory in Learning Analytics in Formal and Workplace Settings. Journal of Learning Analytics, 2(2), 1–4. https://doi.org/10.18608/jla.2015.22.1
Fernando, C., Banarse, D., Michalewski, H., Osindero, S., Rocktäschel, T., & Deepmind, G. (2023). Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution. https://arxiv.org/abs/2309.16797v1
Li, S., Chen, J., Shen, Y., Chen, Z., Zhang, X., Li, Z., Wang, H., Qian, J., Peng, B., Mao, Y., Chen, W., & Yan, X. (2022). Explanations from Large Language Models Make Small Reasoners Better. https://arxiv.org/abs/2210.06726v1
Molenaar, I. (2022). Towards hybrid human-AI learning technologies. European Journal of Education, 57(4), 632–645. https://doi.org/10.1111/EJED.12527

Profile

We expect from applicants:

• a Master degree in Artificial Intelligence, Computational Linguistics,  Computer Science or  related discipline (minimum cum laude)
• strong programming skills
• the ability to do independent research
• a good background and interest in language models
• strong commitment and the ability to work in a team
• a high level of proficiency in English, both spoken and written.
• proficiency in Dutch is a strong plus.

Offer

The position comes with full funding for four years. The funding is available from October 1st 2024. The actual starting date can be negotiated.
The research will be carried out at the faculty of Engineering Science of KU Leuven, campus Heverlee and the Faculty of Arts of KU Leuven, the liguistics research unit, campus Leuven.

Interested?

For more information please contact Prof. dr. ir. Tinne De Laet, tel.: +32 16 32 70 75, mail: tinne.delaet@kuleuven.be or Prof. dr. Tim Van de Cruys, tel.: +32 16 32 35 19, mail: tim.vandecruys@kuleuven.be.

KU Leuven strives for an inclusive, respectful and socially safe environment. We embrace diversity among individuals and groups as an asset. Open dialogue and differences in perspective are essential for an ambitious research and educational environment. In our commitment to equal opportunity, we recognize the consequences of historical inequalities. We do not accept any form of discrimination based on, but not limited to, gender identity and expression, sexual orientation, age, ethnic or national background, skin colour, religious and philosophical diversity, neurodivergence, employment disability, health, or socioeconomic status. For questions about accessibility or support offered, we are happy to assist you at this email address.