WorldCast: Exploring Intelligent Visual Generation in World State Prediction

WorldCast. Exploring Intelligent Visual Generation in World State Prediction. By Wenhao Yan and eleven others, from Tsinghua University, Meituan LongCat, and five other institutions. Published on the fourteenth of October. Its code repository is public and real: sixteen source files.

People take in multimodal observations of the world and forecast how it will evolve. Recent multimodal generative systems have started to show something similar. They can integrate multimodal context and synthesize visual outputs that depict the world changing. To characterise where this is heading, the authors organise model capabilities into a hierarchy of four levels. Doing so, they find that the higher-level tasks remain relatively underexplored, despite all the progress models have made. Lower rungs of that ladder have received attention. The upper rungs have not been explored nearly as much, and that is the gap a new benchmark is meant to fill.

The task they pick to close that gap is multi-step world state prediction. A model is given multimodal observations and diverse references, and has to forecast a succession of future world states, not just the next one. That is demanding in several directions at once. It calls for understanding and generation that are entangled rather than separate, and for extensive world knowledge: enough to turn human action sequences, physical transformations and abstract visual reasoning into visual outputs that hold together.

Put plainly, a model cannot simply understand the scene and then hand off to a separate image generator. Understanding and generation have to work as one, step after step, and each predicted state has to follow coherently from the observations and from the states before it. That is what makes the task high level in the authors' hierarchy.

WorldCast is the benchmark built around that task. It contains three hundred and twenty manually designed cases, spread across eight categories, and the authors use it to evaluate twelve representative models. Every case is designed by hand, and each one asks a model to carry a scene forward through successive states from what it has been shown.

The experiments yield three insights. The first is that frontier models are rapidly pushing the boundaries of intelligent world modeling. The GPT-Image-2.5 System scores seventy-two point eight, and shows early-stage ability to predict visual states on real-world tasks.

The second is that WorldCast exposes substantial gaps between models. Even Nano Banana Pro reaches only forty-two point eight. The best open-source model scores twenty point one. The weaknesses named are handling complex context, dealing with entangled references, and tasks that lean heavily on reasoning.

The third insight concerns where the strength comes from. Strong generative models, the authors report, effectively draw on the capabilities of their underlying language models. Generation quality, on this kind of task, is not only a matter of the image side.

Read together, the three point the same direction. The authors conclude that multimodal generative models need to be capable of dynamic planning and reasoning, and also deeply unified, if they are to simulate real-world states faithfully. The spread of scores is the argument: one system already showing early forecasting ability, Nano Banana Pro far behind it, and open-source models further back still, on a task that sits near the top of the four-level hierarchy the authors drew.