GameVerse: A Minute-Scale Gameplay Dataset for Long-Horizon Interactive World Modeling

GameVerse. A Minute-Scale Gameplay Dataset for Long-Horizon Interactive World Modeling. By Kang He and nine others, from Alaya Lab, the Shanghai Innovation Institute, and two other institutions. Published on the thirty-first of October.

Interactive world models can now generate controllable, dynamic environments from action inputs, but current ones struggle to stay coherent beyond roughly one minute of rollout. The authors place the cause not in model capacity but in a shortage of long-horizon training data. Existing gameplay video datasets mostly hold clips under ten seconds, captioned with coarse text that describes a whole clip without distinguishing what happens within it. That is not enough to learn controllable generation over long horizons. GameVerse is the answer: a minute-scale gameplay video dataset, built specifically for long-horizon interactive world modeling.

It contains sixty-nine point seven thousand clips, spanning one thousand six hundred and fifty-five hours of footage and three hundred and forty-nine point six million frames. The gameplay is diverse, in both first and third person. Of those hours, one thousand three hundred and fifty-eight are continuous two-minute sequences, which the authors describe as an order of magnitude longer than prior datasets.

Every clip carries two kinds of annotation. One is a camera trajectory that stays temporally stable and coherent across minutes. The other is a set of captions that keep static scene context apart from dynamic foreground events, and that line up with event boundaries instead of fixed time windows. A caption marks where something happens, not merely where a clock ticked over.

Clips reach that state through one unified pipeline in four steps: scene-consistent segmentation, quality filtering, camera pose estimation, and event-based caption annotation. The same pipeline is applied to every clip, whether the source is first-person or third-person play, so the whole collection comes out annotated in one consistent way.

The intended users are those training interactive world models who have run into the one-minute wall. A comprehensive analysis reports that GameVerse surpasses existing public datasets in scale, in diversity and in annotation quality. Experiments go further and show that training on it substantially extends the temporal horizon of interactive world models, which is the practical test of whether longer data was really the bottleneck.

The bet underneath this dataset is a diagnosis. If coherence collapses after about a minute because models have never seen what a longer stretch of play looks like, then bigger models will not fix it and longer, better-annotated footage should. GameVerse supplies that footage, and the horizon results are its first evidence that the diagnosis holds.