The field of artificial intelligence is moving towards developing more sophisticated and robust world models, enabling agents to reason about and navigate complex environments. Recent research has focused on improving world models using deep supervision techniques, which have shown to enhance training stability and result in more easily decodable world features. Another direction of research is the development of embodied systems that can understand their own motion dynamics, facilitating efficient skill acquisition and effective planning. Hierarchical reinforcement learning frameworks have also been proposed, allowing for top-down recursive planning via learned subgoals. Noteworthy papers include: Improving World Models using Deep Supervision with Linear Probes, which demonstrates the effectiveness of deep supervision techniques in improving world models. GROVE: A Generalized Reward for Learning Open-Vocabulary Physical Skill, which introduces a generalized reward framework for learning open-vocabulary physical skills. Neural Motion Simulator: Pushing the Limit of World Models in Reinforcement Learning, which presents a world model that predicts the future physical state of an embodied system. Solving Sokoban using Hierarchical Reinforcement Learning with Landmarks, which introduces a novel hierarchical reinforcement learning framework for solving complex combinatorial puzzle games.