GameVerse Dataset Targets World Models' One-Minute Rollout Limit

Interactive world models stall past about a minute of rollout because training clips are too short, not because the models themselves are undersized. That is the argument behind GameVerse, a NeurIPS 2026 dataset posted on arXiv by Kang He, Wenshuo Peng and colleagues, which packages 69.7K gameplay clips spanning 1,655 hours and 349.6 million frames, with 1,358 hours of continuous two-minute sequences.

The paper frames the gap in plain terms. Current models "struggle to maintain coherence beyond roughly one minute of rollout, a limitation rooted not in model capacity but in the scarcity of long-horizon training data," the authors write, noting that existing gameplay video datasets "typically consist of clips under ten seconds, with captions that describe whole clips in coarse and undifferentiated text."

GameVerse's pipeline swaps that recipe. Each clip is paired with a "temporally stable camera trajectory that remains coherent across minutes," and event-based captions split static scene context from dynamic foreground events rather than describing fixed time windows. The team says the set is "an order of magnitude longer than prior datasets" and reports that training on it "substantially extends the temporal horizon of interactive world models."

The abstract stops short of publishing per-model rollout numbers, baseline comparisons, or the specific games sampled, leaving the size of the extension to the full paper.