Compare families of world models.

This is the same reference used in Technical mode. Expand it whenever you need the exact numbers, assumptions or further explanation.

Begin with the output, not the label

Three demonstrations may all be called world models: a system generates convincing pictures, another predicts compact representations, and a third trains a controller using imagined experience. Ask what each predicts, what conditions that prediction, how actions are chosen, and what evidence tests the result.

The lab supplies dated primary-source cards. They are examples for comparison, not a current ranking. Read their actual protocols before transferring a claim into another setting.

Three different research examples

The 2025 DreamerV3 paper describes learned world models supporting behavior learning through imagined trajectories. Its reported control benchmarks do not turn this browser’s small affine cart into a reproduction. A relevant comparison concerns external task outcomes and the resources used to reach them.

The June 2025 V-JEPA 2 paper studies predictive video representations and action-conditioned post-training for planning. A representation prediction objective differs from producing attractive pixels. Video understanding and robotic planning use different outcomes; success on one does not automatically establish the other.

The December 2024 Genie 2 research announcement presents action-conditioned interactive visual environments. Generated observations can support a useful demonstration without establishing exact physics, a task-solving policy or reliable transfer to a physical machine. An announcement and a fully specified experiment also provide different levels of reproducibility detail.

Test a visible counterexample

The four-frame gallery is locally authored, not downloaded output from any named system. A circle starts at coordinate zero. The declared action is −1, 0 or +1, and each interval should change position by 0.2 times that action. For action +1, the expected sequence is 0, 0.2, 0.4, 0.6 scene units.

Disable the consistent rule. The circle remains neatly rendered, yet alternate positions violate the transition contract. Mean absolute position discrepancy measures this particular consistency failure. It is separate from whether the images look pleasing. Changing the action tests whether the sequence responds according to its declared control.

Build a claim and evidence table

For each card record the predicted quantity, observation/action inputs, learning objective, policy role, external evaluator and limitations. Avoid comparing unlike metric values as if they shared a scale. A low latent error may reflect a collapsed code; a plausible rollout may fail action consistency; strong control may rely on a task-specific representation rather than realistic pictures.

Choose an external task, freeze candidate choices before opening outcomes, retain failures, and account for model and environment work. Your annotation is research-reading practice. It becomes an implementation or empirical replication claim only when the corresponding system, protocol and independent evidence actually exist.