World models stop drawing and start obeying: rules, intentions, stable rollouts

Rules before pixels

World-model research is pivoting from rendering video to behaving like a world. Four recent preprints share one theme: models are now being measured and trained on whether they obey rules, not on how they look. For founders building embodied or simulation products, the axis has moved from pixels to rules, intentions, and long-horizon stability. m:…6295

The State World Models paper makes the split literal. Existing world models, the authors argue, often focus on camera movement — letting users explore a generated scene — rather than persistent rules. Their Stateful Harness maintains an evolving world state under specified rules, renders each event with a frozen video model, and uses the generated clip to update the world state for the next turn. To test rule compliance they built WorldState-Bench: 50 counterfactual twin pairs sharing an initial frame and action but differing in hidden state and required outcome. Under rule-only prompting, the best of 11 evaluated video backends produced the required endings for both cases in 32% of pairs; the full harness reaches 92% with the same frozen model. m:…6295

A benchmark that measures dynamics

A second benchmark, AnyGoal, argues the field has been measuring the wrong thing. The goal of a world model, the researchers write, is to capture the dynamics of an environment — yet current models consistently exhibit failures when familiar state and action factors are recombined into unseen combinations. AnyGoal's tasks keep individual state factors and actions within their training support; only the combinations are underrepresented in expert trajectories. The result, across multiple environments and world-model families: consistent failure under recombination. The benchmark ships as an open-source library. m:…4397

Intention-first planning

On the modeling side, a preprint introduces InJePA: intention-first latent planning with coupled JEPA world models. A goal-conditioned JEPA proposes latent intentions, an action-conditioned JEPA predicts future latent states, and the loop closes by feeding predicted consequences back into the next intention proposal and action inference. The point is to plan in latent space through intentions rather than merely evaluating action candidates. As a preprint, the results are author-reported and await replication. m:…3145

Four thousand frames without drift

The long-horizon problem gets a training fix. Streaming video world models trained on short clips drift once rollouts exceed the training horizon, as prediction errors re-enter the context and compound. Fréchet Forcing fine-tunes the model together with its deployed multi-step sampler using self-generated rollouts and real-data statistics, under three constraints the authors call generative anchoring, context fidelity, and reality matching. Training requires no teacher, no online critic, and no long videos. The headline number: fine-tuning MiniWorld-1B on 381-frame clips for at most 36 GPU-hours yields rollouts that remain stable beyond 4,000 frames. m:…6551

What this changes for founders

As reasoning rather than measurement: the moat in world models is moving from bigger video models to rule-faithful, intention-driven, drift-free rollouts — exactly the properties a shipped product needs. But note the evidence gap: AnyGoal shows current models fail at dynamics generalization, and all four of these are single unreplicated reports. Founders should treat rule-compliance benchmarks, not demo footage, as the diligence instrument — and demand third-party replication before betting a roadmap on 4,000-frame stability or 92% rule-following. m:…4397