This series collects my learning notes on world models. I am using it to connect several threads that I keep running into: reinforcement learning, learned dynamics, latent representations, JEPA-style prediction, planning, uncertainty, and model editing or unlearning.

The main question I want to understand is simple to say but hard to answer:

When does a learned model of the world become reliable enough to support decisions, planning, and intervention?

Reading Order

  1. Reinforcement Learning and Abstract World Models

    This is the foundation note. It starts from reinforcement learning and Bellman equations, then moves toward world models, abstract latent-space world models, JEPA-style prediction, and reward hacking. I use this post as the mathematical map before going into specific architectures.

  2. LeWorldModel: A Compact Latent World Model—and Where It Breaks

    This is the first case study. It looks at LeWorldModel as a compact action-conditioned JEPA for planning from pixels, then focuses on the failure modes that matter for reliable control: counterfactual coverage, uncertainty, memory, rollout error, and planners exploiting their own learned model.

Notes To Expand Later

I expect this series to grow around a few recurring questions:

  • how much of a world model should be learned from pixels, latent states, rewards, language, or simulation;
  • how to tell whether a latent representation preserves the state distinctions that planning actually needs;
  • how uncertainty, memory, and counterfactual data coverage change the reliability of model-based control;
  • how to edit, unlearn, or localize a transition belief without damaging unrelated dynamics.

For now, these two posts are the starting point: one for the basic mathematical frame, and one for a concrete latent-world-model case study.