MASS: Multiplayer World Models Simulate 1,024 Concurrent Players
AlayaLab · hf · 2026-08-07
MASS (Multiplayer World Models with Authoritative Shared State) tackles the redundant compute, view inconsistencies, and poor scalability of current video world models in multiplayer environments caused by entangling world state with view-dependent visual latents.
Inspired by multiplayer game architectures, MASS disentangles world dynamics from view rendering:
- Logic Engine: Advances a global, authoritative typed state from joint actions without hand-written transition functions, acting as the sole recurrent memory.
- Rendering Engine: Generates independent and consistent views on demand for any requested camera from the shared state.
This explicit disentangling yields higher state accuracy and lower cross-view inconsistency than state-of-the-art baselines. The model successfully scales to predict worlds with 1,024 concurrent players for 10,000 recurrent steps.
More from Research
- Leanstral: 6B Active Parameters Hits New SOTA in Theorem Proving — AlbertQJiang · 2026-08-07
- The Gray Area of Peer Review: Whose Fault Is It When Papers Get Rejected? — qberthet · 2026-08-07
- ContextMaster Unifies Multi-Shot Video Creation at 16 FPS on a Single GPU — KlingTeam · 2026-08-07
- RSI Over Scale: How Recursive Self-Improvement Could Collapse ASI Costs — imjustnewatai · 2026-08-07
- Rant: The AI Community Loves Reinventing Classic Statistics with New Names — gleech · 2026-08-07
- GitHub Open Source: Build a RAG System from Scratch Using Local LLMs — tom_doerr · 2026-08-07