AI for scientific discovery
Science is a sequential decision problem under uncertainty: which hypothesis to pursue, which experiment to run next, and when to change direction. These are the questions our agents are built to answer.
Why agents
LLMs have absorbed a large part of the scientific literature, and LLM agents can now write code, run simulations and analyse data. But discovery is not a single question and answer. It is a long loop of hypotheses, experiments and revisions, in which each experiment is expensive and most ideas fail. Doing this well needs more than a strong model:
- learning from each experiment: updating beliefs and strategy from results, which is exactly in-context reinforcement learning;
- remembering what was tried and why it failed, across long projects, which is the role of memory systems;
- deciding where to look next: exploration versus exploitation under a limited budget of experiments, and planning several steps ahead;
- reliability: knowing when a result can be trusted and when the agent is fooling itself.
What we bring
My group combines expertise in sequential decision-making, look-ahead search, reinforcement learning with learned models, and LLM agents. We know how to build agents that plan with imperfect information and learn from few interactions, and how to evaluate them rigorously.
Partners wanted
This is a growing direction. I am looking for scientists with a concrete discovery problem, e.g. in chemistry, materials, biology or physics, who have simulators, data or automated labs and want to explore how learning agents can speed up their research. Joint Ph.D. topics and EU project proposals are both welcome. Get in touch.