A new world model design fixes a blind spot in how AI systems plan actions, not just predict what happens next.
Researchers built AD-WM, a joint-embedding world model that adds an action-recovery objective on top of standard latent dynamics, forcing the model to preserve information about which action was taken, not just what state results. On OGBench-Cube, a simulated benchmark, that change lifted hard-start task success from 3.7% to 52.0% over a matched baseline, and outperformed the baseline in four of five test environments overall. The team also paired a frozen V-JEPA 2 encoder with matched DROID post-training and tested the system on a real Franka robot arm. Pick-and-place success there rose from 42.2% to 71.1%, with no extra tuning for that specific lab setup.
The interesting part is not the accuracy bump, it's the diagnosis. The team found that a model's raw prediction error, and even its ranking of every possible action, did not track which model actually worked best in closed-loop control. Only a planning-specific metric lined up with real success, which suggests the field has been optimizing world models for the wrong target.
That's a notable admission for a subfield that has leaned hard on prediction accuracy as its main benchmark. Whether this action-discriminative approach holds up outside one robot arm and one lab is still an open question, but it's a useful reminder that a model can ace the wrong test.