Project: Constraining Large Language Models for Trustworthy and Effective AI Tutors

2024-11-01 – 2028-10-31

Abstract

The progress and performance of Large Language Models and generative AI have accelerated rapidly in the past few years, even being embraced by the general public. However, these systems still have significant weaknesses that temper the excitement, including opaque reasoning behind their outputs, fabricating unsupported information, and ethical concerns around bias, misuse, and their effects on society. Rolling out these models in real-life applications, where unpredictable or incorrect output is unacceptable, requires the study of ways to have fine-grained control over what these models generate. The best-performing models are generally closed source and their usage is often a paid service, while privacy of user data is not guaranteed and the model’s architecture and weights are not known — so access might be lost at any time. Therefore, this research project focuses on improving local models. In applications where precision rather than creativity is important and where ownership of the model cannot be trusted to overseas corporations, there is a pressing need for models over which the application builder has control. This research project tries to solve this by applying, combining and building methods to gain control over generative AI models, to use them in a safe and trustworthy manner. The project evaluates the generative models in a second language tutoring application using social robots, an ambitious use case which stress tests generative AI in real-world environment.