LG AI Research Highlights World Model Research at CVPR 2026

LG AI Research categorized the World Model research that gained attention at CVPR 2026 along two axes: representation methods and usage purposes. Recently, World Models often refer to models that generate future videos based on actions or create playable environments, like the Genie series, but the term is used in a broader sense in Model-Based Reinforcement Learning and Robot Learning.

The targets predicted by World Models are not limited to future videos.

• Action-Controllable Video that generates future Observations
• Predictive Feature representing Semantic Change
• Task-Space Representation such as Object Trajectory, 3D Trace, and Point Flow
• Physical state and Dynamics of the scene

The applications of predicted information are also diverse. Generated futures can be used for Interactive Simulation, feature prediction for Representation Learning, and 3D motion and physical state prediction for Robot Planning and Control. In environments with incomplete observations, one can infer the current hidden state to utilize it for exploration and decision-making rather than directly rolling out the future.

Align While Search (AWS), presented by LG AI Research at CVPR 2026, is an example of such Belief-Space Inference. AWS is not a World Model that generates future Frames or learns Transition Dynamics, but it updates Beliefs about the external world based on observation history in partially observable environments and selects exploration actions to reduce uncertainty.

The article concludes by discussing which World Representation would be useful for the Electronic Design Automation (EDA) tools currently being developed by LG AI Research. The core question is not how completely the world must be generated, but what states should be defined and how they should be modeled for good decision-making.