DWM Decomposes World Effects in Latent Models.

Latent world models are fundamental to modern model-based control, yet their current action-conditioned formulations often conflate different sources of state change. They supervise the next-latent transition with a single target, forcing the model to absorb both action-driven changes and action-invariant "world effects" (intrinsic environmental dynamics like gravity or inertia) into one monolithic learning signal. This entanglement prevents the model from accurately attributing changes to their root causes and hinders the transferability of learned dynamics. Researchers introduce DWM (Decomposed World Model), a supervision-level framework designed to operationalize this decomposition. DWM augments a latent world model's predictor with an auxiliary "world head," which is regularized to be action-invariant through a normalized world-contrastive objective. The original prediction head is then coupled to this world head via an orthogonality constraint. Together, these components induce an explicit additive decomposition of the predicted transition into distinct action-invariant and action-driven components, all without altering the underlying architecture or inference pipeline. To rigorously evaluate DWM, new "W-variants" of three standard control benchmarks (PushT-W, Reacher-W, TwoRoom-W) were created, each incorporating distinct action-invariant dynamics. DWM not only matched strong baselines on the original benchmarks but also delivered a significant mean absolute improvement in CEM planning success across the new W-variants, demonstrating its effectiveness in environments with persistent world effects.