Bagel Labs is an AI research lab that trains frontier diffusion models across commodity hardware instead of one uniform GPU cluster. Our method, Distributed Diffusion Models or DDM, trains many smaller expert models independently with no gradient synchronization, and a lightweight router combines them at inference. Paris-1 proved the idea on images and Paris-2 proved it on video, where three 11B experts and a router beat a monolithic baseline trained on the same compute by more than 50% on FVD. We are now applying DDM to physical AI, where world models, action, and simulation are still wide open.
We ignore years of experience and pedigree. If you have research taste and a real idea about how generative models, action, or world state should work, we want to hear from you. Every requirement below is flexible for someone with the depth to back it up.
Role Overview
The role is deliberately broad and much of the science is still unsettled. We want exceptional researchers working on diffusion and flow models, world models, latent dynamics, action representations, simulation, and embodied generalization. You will help decide which directions are worth chasing as we show that DDM matters for physical AI, moving from public benchmarks toward specialization, compositional generalization, and the prediction of action and world state.
What You'll Do
- Set and pursue your own research directions across diffusion, world models, action representations, latent dynamics, and embodied policy learning.
- Push DDM further for physical AI, including distributed world models, expert ensembles, routing, specialization, and compositional generalization.
- Explore latent action representations such as an action RAE as a reusable layer for embodied AI.
- Design clean experiments on physical AI benchmarks with honest baselines, metrics, and failure analysis.
- Make your results legible to both researchers and the infrastructure team that builds on them.
- Work with systems engineers so your experiments stay reproducible and scale across heterogeneous compute.
Who You Might Be
You are a researcher with strong taste for problems in physical systems, maybe in diffusion or flow models, world models, video generation, simulation, representation learning, or imitation learning and RL. You design experiments that answer real questions, and you can tell a promising result from a fragile one. You do not need to have worked on our exact stack. Adjacent expertise counts for a lot if it comes with a real idea about action, world state, embodiment, or distributed training.
Desired Skills
- Strong research judgment in one or more of diffusion, flow, and score models, world models, representation learning, robotics, simulation, imitation learning, RL, or video and multimodal generation.
- A feel for clean experiment design, including baselines, metrics, and failure modes.
- Fluency reading and implementing recent ML papers.
- Clear technical writing.
- Strong Python and a modern ML framework.
What We Offer
- Competitive compensation and meaningful equity.
- A deeply technical culture where frontier ideas get debated, tested hard, and built.
- Room to pursue open ended, frontier research and shape the agenda while the lab is still small.
- Paid travel to the top ML conferences around the world.
