Factual next-state prediction can hide the action differences model predictive control needs when it compares candidates from the same state. On Sept. 24, 2026, a Nanjing University team posted AD-WM (arXiv:2609.30264), an action-discriminative joint-embedding world model aimed at that counterfactual gap.
In plain terms, AD-WM is not a rolling video-action denoiser and not a change to the planner itself. It combines residual latent dynamics with predictor-level action-recovery regularization: an inverse-dynamics head plus a normalized recovery objective motivated by conditional mutual information. The auxiliary heads are discarded at test time. CEM-based MPC stays the same.
On OGBench-Cube, authors report hard-start success rising from 3.7% for a matched LeWM reproduction to 52.0% for AD-WM. Mean success improves over that matched baseline in four of five simulation environments. On PushT, authors report a slight drop from 94% to 92%, so the sim scoreboard is not a uniform win.
With a frozen V-JEPA 2 encoder and matched DROID post-training, authors report zero-shot Franka basic pick-and-place success rising from 42.2% for matched V-JEPA 2-AC to 71.1%, without lab-specific adaptation. That transfer uses the authors' single Franka setup and shared deployment stack.
The project page hosts overview and demo videos; GitHub code and Hugging Face model weights are live. All scoreboard figures are author-reported from the paper and project page. Independent reproduction outside the authors' protocol is not yet available in the materials reviewed for this article. alphaXiv indexes the same arXiv record.