Bayesian Generative AI

For better or for worse, Generative AI is changing our world.

A key challenge in Generative AI is many-shot jailbreaking—where a language model, despite being explicitly trained to reject harmful or unethical responses, can be manipulated into producing them when provided with a large enough number of well-crafted “in-context” examples. This raises critical questions: Why do AI models adapt so predictably to in-context demonstrations? What controls their ability to generalize, even when it contradicts their training?

This project will involve developing a theoretical framework to explain in-context learning by using Bayesian learning principles. Among other things, Bayesian learning gives AI systems the ability to quantitatively express a degree of belief about a prediction or statement.

By bridging deep learning theory, Bayesian statistics, and generative modelling, this work will advance our understanding of both the capabilities and vulnerabilities of modern AI systems. This will have potential impact in translational fields, which often aim to use generative AI in a responsible way.

This PhD project is part of a larger cohort of projects in the Monash AI Institute, a recently expanded program which brings together PhD students and interdisciplinary academics throughout Monash who are interested in advancing AI related research. This project is jointly supervised by Dr. Russell Tsuchida (Department of Data Science and AI in the Faculty of Information Technology) and Associate Professor Susan Wei (Department of Econometrics and Business Statistics in the Faculty of Business and Economics).

Please email russell.tsuchida@monash.edu and susan.wei@monash.edu for any questions related to this project.