# General World Action Model

> Bagel Labs is a physical AI research lab building a general world action model for autonomous robot control across tasks, environments, and robot types.

The model is designed to improve recursively through experience. This is the direction of the research program, not a claim that autonomous recursive learning or a combined PARIS–WorldDiT system has already been validated.

## Training architecture: PARIS

PARIS is Bagel Labs' training architecture. It lets different parts of a robotics model learn independently, then brings them together as one.

In separate published tests with data and compute held equal, PARIS 1.0 reported a 24% improvement in image generation and PARIS 2.0 reported a 50% improvement in video generation. These results do not establish robotics performance.

- [PARIS 1.0 paper](https://arxiv.org/abs/2510.03434)
- [PARIS 2.0 paper](https://arxiv.org/abs/2605.26064)

## Model architecture: WorldDiT

WorldDiT is Bagel Labs' compact model architecture for robot learning and control. It sits on the reported Pareto frontier for model size and task success in LIBERO simulation.

This is simulation evidence, not real-robot validation or deployment.

- [WorldDiT paper](https://arxiv.org/abs/2607.23909)
- [WorldDiT open release](https://huggingface.co/bageldotcom/worlddit)

## Bagel Labs

- [Research](https://blog.bagel.com/)
- [Careers](https://www.bagel.com/careers)
- [Hugging Face](https://huggingface.co/bageldotcom)
- [GitHub](https://github.com/bageldotcom)
- [LinkedIn](https://www.linkedin.com/company/bageldotcom)
- [X](https://x.com/bageldotcom)

For detailed terminology and evidence boundaries, see the [full Bagel Labs agent context](https://www.bagel.com/llms-full.txt).