Bagel Labs is an AI research lab and infrastructure company building distributed training systems for diffusion-heavy physical AI. Our work started with the Paris family of image and video models and now extends toward physical-AI workloads where world models, action representations, simulation, and heterogeneous compute become first-order bottlenecks.
We ignore years of experience and pedigree. If you have high agency, strong research taste, and original ideas, we want to hear from you. Every requirement below is flexible for a candidate with enough judgment and technical depth.
Role Overview
You will help define Bagel's physical-AI research program. The role is intentionally broad enough for exceptional researchers across diffusion models, world models, latent dynamics, action representations, simulation, embodied generalization, video/multimodal modeling, and decentralized training.
The concrete proof motion is Paris 3: show that Bagel's DDM approach can matter for physical-AI workloads, starting from physical-AI benchmarks and moving toward specialization, compositionality, and action/world-state prediction.
Key Responsibilities
- Develop research directions across diffusion models, physical-AI world models, action representations, latent dynamics, simulation, embodied policy learning, and video/multimodal generative models.
- Contribute to the Paris 3 proof motion, including decentralized world models, DDM expert ensembles, action/world-state prediction, routing, specialization, and compositional generalization.
- Explore Action RAE / latent action encoder directions as a reusable representation layer for embodied AI.
- Design and run experiments on physical-AI benchmarks and datasets such as LIBERO, CALVIN, DROID, Open X-style robot data, simulation data, or adjacent video/world-model evaluations.
- Compare against relevant baselines and make the results legible to both researchers and infrastructure buyers.
- Work with systems engineers to make research experiments reproducible, observable, and scalable across heterogeneous compute.
First 90 Days
- Pick a high-leverage Paris 3 research thread around decentralized world models, action representations, diffusion policies, world-action prediction, or compositional generalization.
- Produce a working experiment plan with clear baselines, datasets, success metrics, and failure modes.
- Run the first credible comparison and write the internal technical note.
- Help decide whether the result should become a paper, public artifact, customer-evaluation benchmark, or next-scale experiment.
Who You Might Be
You might be a diffusion researcher, robotics/world-model researcher, video generation researcher, simulation researcher, representation-learning researcher, RL/imitation-learning researcher, or applied scientist with strong taste around physical systems.
You do not need to have worked on exactly Bagel's current stack. Adjacent expertise is valuable if it brings a real idea about how generative models, action, world state, embodiment, or decentralized training should work.
Desired Skills
- Strong research judgment in one or more of diffusion, flow/score models, world models, representation learning, robotics, simulation, imitation learning, RL, video generation, or multimodal modeling.
- Ability to design clean experiments and reason about baselines, metrics, and failure modes.
- Comfort reading and implementing recent ML papers.
- Clear technical writing.
- Strong Python and modern ML framework experience.
What We Offer
- Top-of-market compensation and meaningful equity.
- A deeply technical culture where frontier ideas are debated, stress-tested, and built.
- Time and support to pursue open-ended research with a concrete company proof motion.
- Flexible location for the right candidate, including relocation support when appropriate.
- Paid travel opportunities to top ML conferences around the world.