'From Pixels to States' Is HuggingFace's Most-Upvoted Paper This Week at 407 — Reframing World Models as Game Engines

arXiv / HuggingFace Daily Papers·medium signal

Alaya Studio's paper (arXiv 2607.14076) argues interactive world models should be built and evaluated as game engines operating over explicit state, not as pixel-space video predictors — and pulled 407 upvotes, roughly 2.3x the next-highest paper on the July 17 feed. The engagement gap is the story: after a year of video-diffusion world models (and Nvidia's Cosmos 3 Edge push on July 17), the research community is voting loudly for state-based simulation as the tractable path.