Joint World Model And Policy Training With Shared Latent Representation

* ⏫ 02/2023: [Grounding Large Language Models in Interactive Environments with Online RL (GLAM)](https://arxiv.org/abs/2302.02662)

Unique: DreamerV3 uses a unified loss function that jointly optimizes reconstruction, dynamics, and policy objectives with learnable loss weights, enabling the policy to guide world model learning. This contrasts with prior approaches (PlaNet, Dreamer v1/v2) that trained world models and policies sequentially or with fixed loss weight ratios.

Achieves better sample efficiency than sequential training by having the policy guide world model learning toward control-relevant features, while maintaining stability through careful loss balancing and shared representation learning.