GameVerse released 1,655 hours of gameplay video including 1,358 hours of continuous two-minute sequences, an order of magnitude longer than prior datasets
AspireAI community · arXiv
*Interactive world models can generate controllable environments from actions, but they lose coherence after roughly a minute - and the reason turns out to be data, not capacity.* Existing gameplay video datasets are mostly clips under ten seconds with captions that describe the whole clip in one undifferentiated block, which is not enough to learn a controllable long horizon. GameVerse is built for exactly that: 69.7 thousand clips spanning 1,655 hours and 349.6 million frames across first- and third-person gameplay, including 1,358 hours of continuous two-minute sequences. Every clip runs through scene-consistent segmentation, quality filtering, camera pose estimation and event-based caption annotation, and each is paired with a camera trajectory that stays coherent across minutes plus captions that separate static scene context from dynamic foreground events and align to event boundaries rather than fixed time windows. Training on it substantially extends the temporal horizon of interactive world models. The limitation: the authors report the horizon extension qualitatively rather than as a single headline number, and the data is gameplay, so transfer to other embodied domains is not established. Accepted at NeurIPS 2026.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.