Heejeong Nam

Abstract

Humans do not memorize every scene or transition to learn. Instead, we form reusable concepts and knowledge from observation, such as how actions move an agent in different directions or how rigid objects can interact and push one another.

Once such knowledge is learned, it becomes much easier to reason about a slightly different goal or to think about counterfactual situations without relearning the environment from scratch. This raises a natural question: can world models do the same?

This talk focuses on two related questions: whether world models truly learn the underlying dynamics of the environment, and how models can be encouraged to learn dynamics that are more reusable and generalizable.

Abstract

World models are increasingly moving away from reconstruction and toward prediction in latent space. In this talk, we will present two recent JEPA-based approaches that illustrate this shift from complementary angles. Link to slides.

Causal-JEPA induces object-level relational bias to promote representations that capture entities, and interactions, leading to stronger reasoning and more efficient planning. LeWorldModel shows that such predictive world models can also be trained stably end-to-end from raw pixels using a minimal objective and a clean architectural recipe, while remaining competitive on control tasks. Taken together, these works argue for a unified view of world modeling: predictive latent learning becomes most powerful when combined with both structural bias and architectural simplicity. This perspective suggests a promising path toward robust world models that support abstraction, reasoning, and control.

Abstract

Humans can understand complex changes in a visual scene by recognizing simpler, reusable patterns of motion and interaction. For example, even when objects look different or appear in unfamiliar situations, we can recognize similar ways in which they move and affect one another.

This suggests that understanding visual dynamics may require more than predicting entire transitions. Instead, models may benefit from learning reusable components of change that can be recombined across different situations.

This talk explores how models can discover such reusable motion primitives directly from videos, and how these primitives can serve as building blocks for learning more generalizable representations of actions and visual dynamics.

Title: Toward Practical Causal Learning in the Real World

Hosted by: Sangwoo Mo, Assistant Professor at Pohang University of Science and Technology (POSTECH)

Slides

Abstract

Causal learning aims to uncover the underlying mechanisms of the world, rather than merely capturing correlations in observed data. However, learning identifiable causal representations often requires strong assumptions, and it remains unclear how these theoretical insights can benefit practical learning systems.

World models offer a promising setting for exploring this connection, as understanding physical dynamics requires reasoning about how objects interact and influence one another.

This talk explores how causal principles can inspire more structured and efficient visual world models, focusing on learning interaction-dependent dynamics through object-level latent interventions. It also discusses the challenges of bridging causal representation learning and practical world modeling.