For students

I supervise M.Sc. theses and student research projects on prompt optimization and multi-agent systems. Ideally you come with a question of your own in this area. If you would rather start from one of mine, these are the three I would most like to see someone take on right now.

1. Meta-prompt optimization Prompt optimizers such as CAPO are themselves driven by prompts: a meta-prompt tells the model how to mutate and combine candidates. So far that meta-prompt is a fixed part of the method. The question is whether optimizing it makes the optimizer better, and by how much.
2. Scaling laws in agentic systems For a single model we know roughly how performance grows with parameters and data. For a system of agents there are no such curves yet: what happens when you add agents, rounds of interaction, or tokens per agent? The thesis would measure those curves on a fixed set of tasks and find where the gains stop.
3. Agentic systems for robotic control A robot gives an agent system a task where the environment answers back: the agents plan, act, and correct from what the robot senses, and a wrong plan shows up as a wrong movement. The thesis would build such a system, start in a simulator, and move to a real robot if it works.