Research Scientist - World Model

You'll turn Luma's industry-leading generative video models into world models: interactive, controllable, physically faithful, and useful as a substrate for embodied reasoning. This is the role at the center of the thesis. You'll invent next-generation world-model architectures and the controllability that lets an agent step into a generated world, and own the metrics that define success.

It fits a researcher with deep generative-modeling or model-based-RL expertise who has trained models to the limits of a multi-node cluster. If you want a narrow, well-scoped research problem, this is broader and more open-ended than that. What You'll Own - Invent next-generation world-model architectures (diffusion, transformer, autoregressive, or hybrid), focused on controllability and physical consistency. - Develop controllability mechanisms - action conditioning, view conditioning, long-horizon rollouts - that let an agent step into the world. - Define and own the metrics: physical fidelity, long-horizon coherence, action-following, and downstream usefulness for policy training. - Run scaling studies that show where compute, data, and architecture pay off. - Publish at the frontier and contribute to the open-source release that is the long-term deliverable.

First 90 Days One way the first 90 could unfold. - Days 1-30 - Immerse & Diagnose: Get deep on the current video models and where they fall short as world models. - Days 30-60 - Ship & Validate: Prototype a controllability mechanism or architecture change and measure it against physical-fidelity and action-following metrics. - Days 60-90 - Scale & Systemize: Run scaling studies and push the most promising direction toward the open release.

What You Bring - PhD or equivalent research record in ML, computer vision, robotics, or a related field. - Deep expertise in at least one of: large-scale generative modeling (video/3D/world), self-supervised representation learning, or model-based RL. - Strong PyTorch and large-scale training experience, to the limits of a multi-node cluster. - A research record the field knows (top-venue publications and/or widely used open releases).

Nice to Have - Prior work on world models, model-based RL, generative video, neural simulation, or 4D scene representations. - Experience using generative models for downstream embodied tasks (planning, control, evaluation). - Enthusiasm for open-sourcing frontier models. About Luma: Luma's mission is to build unified general intelligence that can generate, understand, and operate in the physical world.

We believe multimodality is critical for intelligence - the next step beyond language models comes from vision. Luma is an equal opportunity employer.