Opportunities

These are open projects within ATOM, led by me or by colleagues, for students and researchers who want to join. All projects are remote by default, and in select cases can be hosted in person in Tübingen (see the FAQ).

Open projects

Principled hierarchies

Universal world modeling

How should agents learn abstract, hierarchical models of their world? This project connects hierarchical reinforcement learning, causal models, multiscale physics and hierarchical active inference, in theory and in practice.

Who we’re looking for and how to apply

We are looking for: a student with a strong theoretical background and an interest in machine learning; relevant experience is a plus. Ideally full-time or close to it. You would work with a full-time PhD student on the project and experts across ATOM, and lead a paper as first or joint first author, targeting submission by May 2027.

Apply: , cc’ing , , and , with the subject “[Principled hierarchies] Your name”, your CV and one or two paragraphs on why you are interested, which areas interest you most, how many hours per week you can commit, and anything not obvious from your CV.

Natural building blocks in practice

Universal world modeling

Develop learnable world-model building blocks that work on their own and combine into hierarchical active inference agents.

Who we’re looking for and how to apply

We are looking for: up to two contributors, one leaning towards modeling and learning, the other towards benchmarking and implementation, with strong Python and a working knowledge of probabilistic models; relevant experience is a plus. Ideally full-time or close to it. The goal is a manuscript and open-source code by May 2027.

Apply: , cc’ing , with the subject “[Natural building blocks] Your name”, your CV and a few paragraphs on why you are interested, which strand you lean towards, and how many hours per week you can commit.

LLM-driven structure learning

Model discovery

Use large language models to propose and refine the structure of world models for active inference agents, keeping an interpretable trace of every change. An existing pipeline shows promising first results, and many scientific questions remain open.

Who we’re looking for and how to apply

We are looking for: a motivated student or a team of two to take the project forward, with solid Python and an interest in LLM-based program and model synthesis; experience with probabilistic modelling or active inference is a plus. Ideally full-time or close to it, although strong relevant experience can make up for fewer hours. The goal is a paper at ICML (deadline expected in late January 2027) or NeurIPS (expected in mid-May 2027).

Apply: , cc’ing , with the subject “[LLM structure learning] Your name”, your CV, a few lines on why this project interests you, and how many hours per week you can commit.

Benchmarking ATOM agents

Evaluation

What are agents that learn like we do good at, and how should we measure it? Review existing reinforcement learning and world modeling benchmarks, learn about how children learn, and design new evaluations.

Who we’re looking for and how to apply

We are looking for: a student interested in reinforcement learning, world modeling and evaluation, and ideally in how children learn; relevant experience is a plus. Ideally full-time or close to it. The goal is a paper at ICML (deadline expected in late January 2027) or NeurIPS (expected in mid-May 2027).

Apply: with the subject “[Benchmarking ATOM agents] Your name”, your CV, a few lines on why this project interests you, and how many hours per week you can commit.

Propose a project

Have an idea of your own? Students and researchers alike are welcome to propose a project: with the subject “[Project proposal] Your name”, your CV and a short description of the project you have in mind.

Frequently asked questions

Is this a remote or in-person position?

Remote by default. Hosting in person in Bernhard Schölkopf’s Empirical Inference Department at the Max Planck Institute for Intelligent Systems is occasionally possible, depending on capacity and timing.

Is this a paid position?

Unpaid by default. Paid positions in Bernhard Schölkopf’s Empirical Inference Department at the Max Planck Institute for Intelligent Systems are occasionally possible, depending on funding, capacity and timing.

Can I do my master’s thesis on one of these projects?

Yes: apply to the project(s) that interest you and say you would like to do it as your master’s thesis (ETH and Tübingen students can be co-supervised with Bernhard Schölkopf). We will get back to you if there is a fit.

Any further questions?

Drop me a quick note by with the subject “[Supervision Question] XXX”.

Past projects

Past projects have been a team effort across colleagues, and every master’s and PhD student who has completed a project has published from it. Papers have appeared at NeurIPS, at workshops at NeurIPS and ICML, and as an oral at the International Workshop on Active Inference, most often first-authored by the students who led them. To date: 11 students supervised, 6 publications and a best poster award.

Papers from past projects (students at the time in bold)

The complete list of supervised students is on my CV*.

*Updated sparingly.