WorldDiT: A Unified Architecture for World and Action Modeling

WorldDiT is a pareto frontier robot world-action model architecture without a VLM backbone

We’re releasing WorldDiT, a pareto frontier model architecture bringing world and robot action modeling into one compact diffusion transformer, without relying on a large pretrained VLM action backbone.

WorldDiT in action

Three reasons WorldDiT is different

• At 399 million total parameters, WorldDiT lies on the reported LIBERO Pareto frontier for model size and mean success.

• Among publicly released methods in the comparison that do not use a large pretrained VLM action backbone, it reports the highest mean success.

• WorldDiT unifies future visual world state prediction with continuous robot action generation in a single diffusion transformer.

Reported LIBERO success against total model parameters for 24 methods.

🤗 Run WorldDiT from Hugging Face with the released checkpoints and complete inference runtime.

📄 Learn more about the architecture, training, and results in our technical report on arXiv.

If you use WorldDiT in your research, feel free to cite the paper.

@article{260723909,
  title={{WorldDiT: A Unified Diffusion Architecture for World and Action Modeling}},
  author={Sen Wang and R. Gnana Praveen and Bidhan Roy and Marcos Villagra},
  year={{2026}},
  eprint={{2607.23909}},
  archivePrefix={{arXiv}}
}