World Superintelligence Labs
Build world models with superintelligence, step by step.
The next step beyond language intelligence: world models that learn from multimodal world data, simulate action-conditioned futures, and turn that simulation into planning in the physical world.
Thesis
Language models showed how far intelligence can go when trained on text. WSI asks what comes next when the training signal becomes the world itself.
Roads, factories, homes, weather, bodies, machines, and robots do not arrive as clean tokens. They arrive as continuous, noisy, multimodal streams where small actions change what can happen next.
WSI Labs builds world models for that setting. The model must read multimodal evidence from the world, keep the causal structure that matters, and generate future states or observations conditioned on action.
This is where perception becomes planning. A driver, robot, or embodied agent should not only recognize the scene in front of it; it should evaluate what the scene becomes under a turn, a brake, a reach, or a wait.
We measure progress in closed-loop behavior: better predictions, safer plans, and representations that make physical knowledge usable for action.
Research
World models that understand multimodal streams, generate action-conditioned futures, and hold up in demanding physical systems.
- Physical world models: learning compact latent states that keep the causal and geometric structure needed to generate action-conditioned futures.
- Map-free intelligent driving: building driving intelligence that understands scenes, intent, uncertainty, and local physics without depending on brittle high-definition maps.
- Robots that reason in motion: connecting perception, memory, and action so robots can recover from change, choose useful experiments, and work in real environments.
- High-dimensional planning: developing reasoning systems that can keep many constraints in view, simulate consequences, and select safe action sequences.
Recent Work
Research notes from the lab: what changed in the model, what was measured, and why it matters for physical AI.
- ReWorld: Better Representations for World Action Models. arXiv 2026, 2026.06. A short note on directly supervising intermediate world and action representations for better generation quality and safer planning. world action modelsplanningsafety
- DriveLaW: Planning from a Latent Driving World. CVPR 2026, 2026.06. A short note on connecting driving video generation and trajectory planning through a shared latent world representation. latent world modelsdrivingvideo generation
Contact
If you are building physical AI systems, robotics platforms, autonomous driving stacks, or new model architectures, we would like to talk.