Causal World Models

Student selection is now closed! We have informed those who were accepted. We received over 150 applications, many of which included good reasons to join, but we can only accept a limited number of students because our resources are also limited!

Course Information

• Organized by:  Prof. Joschka Boedecker, Julien Brosseit

• Kickoff meeting: Thursday 16.10.2025, 1 p.m. - 1:30 p.m. Nexus Lab

• Location: Intelligent Machine-Brain Interface Technology (IMBIT), Nexus Lab 
• HISinOne: 11LE13S-7362-MB - Causal World Models (Seminar)
• Language: English
• Email:cwm-seminar@informatik.uni-freiburg.de

Overview

World Models (Ha et al., 2018) are learned simulators used by agents to infer knowlegde, plan and make informed decisions in their environment. They are proving to be an effective approach in reinforcement learning (RL), robotics, generative AI, and other areas of machine learning. After all, they share similarities to how we humans think about our environment. 

While standard approaches to learning world models are able to capture correlations in the data, they may fail under interventions and distribution shifts. Causal World Models (CWM) (Li et al., 2020) are able to capture underlying factors to answer counterfactually queries such as "What would have happened if I had acted differently?" leading to improved generalization, explainability and robustness.

In this seminar, we will take a closer look at CWMs, about their inner workings, and why they are so important for building effective World Models.

Format

The course will be given in person, in the form of a block seminar, where papers are read and presented by students in the form of a scientific poster.

Paper Voting Process

Students can rank their preferences in our ILIAS course.

Timeline

Paper List

1. "Causal World Models by Unsupervised Deconfounding of Physical Dynamics" (Li et al., 2020) [link]
2. "SPARTAN: A Sparse Transformer Learning Local Causation" (Lei et al., 2024) [link]
3. "A Causal World Model Underlying Next Token Prediction: Exploring GPT in a Controlled Environment" (Rohekar et al., 2024) [link]
4. "Towards Generalizable Reinforcement Learning via Causality-Guided Self-Adaptive Representations" (Yang et al., 2024) [link]
5. "Decision Transformer: Reinforcement Learning via Sequence Modeling" (Chen et al., 2021) [link]
6. "Marrying Causal Representation Learning with Dynamical Systems for Science" (Yao et al., 2024) [link]
7. "Training Agents Inside of Scalable World Models" (Hafner et al., 2025) [link]
8. "Causal-oriented representation learning for time-series forecasting based on the spatiotemporal information transformation" (Cai et al., 2025) [link] 
9. "Variational Causal Dynamics: Discovering Modular World Models from Interventions" (Lei et al., 2022) [link]
10. "Learning World Models with Identifiable Factorization" (Liu et al., 2023) [link]
11. "Generative Emergent Communication: Large Language Model is a Collective World Model" (Taniguchi et al., 2025) [link]
12. "Robust agents learn causal world models" (Richens et al., 2024) [link]
13. "CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models" (Zhao et al., 2025) [link]
14.  "The Predictron: End-To-End Learning and Planning" (Silver et al., 2017) [link]
15. "Learning Particle Dynamics for Manipulating Rigid Bodies, Deformable Objects, and Fluids" (Li et al., 2019) [link]
16. "Causal Discovery in Physical Systems from Videos" (Li et al., 2020) [link]
17. "Better Decisions through the Right Causal World Model" (Dillies et al., 2025) [link]
18.  "An Efficient Dialogue Policy Agent with Model-Based Causal Reinforcement Learning" (Xu et al., 2025) [link]
19. "AVID: Adapting Video Diffusion Models to World Models" (Rigter et al., 2024) [link]

Resources

Poster guideline as pptx or pdf