World models that predict the future without collapsing into a single boring guess just got a simpler way to avoid that trap.
A paper titled "No Gaussian Required: Contrastive Inverse Dynamics for JEPA World Models" (arXiv:2608.17542) tackles a known weak spot in Joint-Embedding Predictive Architectures, or JEPAs, which learn by predicting future embeddings rather than raw pixels. That setup has a shortcut: the encoder can output a constant vector for everything and technically satisfy the training objective while learning nothing. Existing systems, including LeWorldModel, block this with a regularizer called SIGReg that forces internal representations to match an isotropic Gaussian distribution. The paper's method, AC-MTM (Action-Contrastive Masked Transition Modeling), instead adds a training-only head that must identify which action, out of a batch of candidates, produced a given state transition - a task a collapsed encoder cannot pass. That head is discarded after training, so the deployed model runs identically to LeWorldModel.
On four standard pixel-control benchmarks, AC-MTM matched SIGReg's performance from scratch. On a harder task, OGBench's multi-object Visual Scene benchmark, it did notably better: 80.0% success versus 58.0% for SIGReg, a 20-24 point gap that held across every training seed, against a 52% random-policy baseline.
That gap is the real finding. Forcing representations into a predetermined Gaussian shape is a modeling assumption, and assumptions carry a cost when the environment doesn't match them. Multi-object scenes are messier than single-object pixel tasks, and that's exactly where the prescribed geometry looks like a bottleneck rather than a stabilizer. AC-MTM also drops the target networks, stop-gradients, and pretrained encoders that similar methods often lean on.
One harder benchmark isn't proof this generalizes to real robots or messier footage, and the authors flag their own action-space assumptions as a limit. The code is public on GitHub for anyone who wants to check the numbers.