Advanced topics in causality and causal ML research

General advice

The plan is for each group to choose one paper, read it carefully and reproduce the experiments in the original paper (or at least as much as possible), and propose some small extension. I will be helping with advice and evaluate which extensions are feasible within the course timeframe.

Possible ideas:

• A simple (or not, depending on the ambition) possible extension is to apply the methods to new (ideally real-world) datasets or problem
• Adapting the methods to new settings beyond their current theory or relaxing assumptions at least empirically is always interesting
• Creative combinations of methods are also welcome

Caveats:

• I just moved to UdS so I'm not exactly sure about the computational resources available to you. Once we figure out what are you interested in, I will investigate, but in any case, there are always projects that can be run on laptops.
• I have selected some recent and exciting papers that (afaik) have reasonable or well-documented codebases. In most cases, I know the contributors, so we can ask for help if needed. In principle, you can also propose something else, but this means I might not be able to give you the same level of advice.

Causal discovery

Methods to improve the scalability of causal discovery:

• Sequential Non-Ancestor Pruning - SNAP [Schubert et al. 2025] - project page with paper and code
• Local Optimal Adjustments discovery - LOAD [Schubert et al. 2026] - project page with paper and code

Methods to improve accuracy of causal discovery by integrating background knowledge, either through imperfect experts (e.g. LLMs) or through iterative refinement:

• Imperfect experts: Guess2Graph [Hiramath et al. 2026] - project page with paper and code
• Expert-in-the-loop [Ankan et al. 2025] - paper, example notebook

Datasets that might be interesting to try:

• Causal graph of an assembly line from Bosch [Göbler et al. 2024]: https://github.com/boschresearch/causalAssembly

Causal representation learning (CRL)

Methods that can disentangle causal variables from realistic images based on interventions or actions:

• Interventional CRL methods: CITRIS [Lippe et al. 2022] and iCITRIS [Lippe et al. 2023] - project page with papers and code, example notebook with pretrained models
• CRL based on actions: BISCUIT [Lippe et al. 2023] - project page with paper, code, pretrained model and datasets

Methods that do not require actions, but instead assume multiple simultaneous views

• Multiview-CRL [Yao et al. 2024] - project page with paper and code

 

Downstream tasks / Causality-inspired ML/RL

Methods that exploit causality or CRL for guarantees in interpretability or XAI:

• Guarantees for Concept Bottleneck Models: [Fokkema et al. 2025] paper, code
• Interpretability of LLMs through mixtures of causal models: [Pislar et al. 2025] paper, code

Generalization in RL through causal representations:

• AdaRL: What, Where, and How to Adapt in Transfer Reinforcement Learning [Huang et al. 2022] project page with paper and code
• Causality-guided Self-adaptive Representation-based approach - CSR [Yang et al. 2025] project page with paper and code

Using causality to improve bandits:

• (a bit simpler) Structural causal bandits [Lee et al. 2018] project page with code and paper
• (extension of the previous one) Non-stationary causal bandits [Kwon et al. 2025] project page with code and paper