EmergingMultimodal and generative world models

World Model Researcher

Also called Predictive World-Model Scientist

A World Model Researcher builds models that predict how an environment will change, given what has happened and what an agent does next. That gives the company a learned simulator for planning, testing and training agents.

What they do

Designs and trains models that learn how environments behave from video, sound, text and action data. Chooses what the model should represent, and what it can safely ignore. Builds rollout tests that run the model forward for many steps and compare the result with trajectories it has never seen. Looks at where predictions drift, blur or break physical sense, then changes the architecture, objective or data mix. Works with planning and agent teams to check that the model helps them make better decisions. Writes up findings so others can reproduce them.

What a strong hire delivers

Within the first months, a strong hire delivers a working baseline world model with a clear evaluation set of unseen trajectories. They show how prediction error grows over longer horizons and where it fails. They agree a measure of planning usefulness with the downstream team. They leave behind reproducible training and evaluation code, and a short list of the next experiments worth running.

How you can tell they are good

Ask them to walk through a world model they built and how they tested it. Strong people explain their latent representation, why they chose it, and how they separated real dynamics from memorised training data. They talk about long-horizon drift without prompting. They can show a failure case and what they changed. Be wary of candidates who only show sharp-looking samples and have no held-out rollout evidence.

Why it is hard to fill

The title barely exists yet. The work is usually done under labels such as multimodal researcher, video generation researcher or simulation scientist. Few people have trained these models at scale and also evaluated them against planning use. Candidates often come from one side only: generative media with weak dynamics rigour, or control and simulation with little experience of large multimodal training. Expect to assess skills, not titles.

Where to find them

Look in multimodal and video generation teams, robotics and embodied AI groups, and simulation or digital twin research. Reinforcement learning researchers who worked on model-based methods are a strong pool. Many sit in frontier labs, well-funded research startups, autonomous systems companies and university labs. Their current titles are often research scientist or research engineer, so read their papers and projects, not their job titles.

Indicative pay

Not enough published evidence yet to give a fair range for this role, which is common for work this new. We benchmark it for you as part of a search.

Interview questions for a World Model Researcher

  1. Tell us about a world model you trained. What did it represent, and how did you test it on trajectories it had not seen?

    A strong answer names the data, the latent design and the held-out split, and explains how leakage was avoided. It reports honest failures. Worry if evaluation was mostly visual inspection of samples or if the test set resembled the training set.

  2. How do you measure prediction error over long horizons, and why is a single-step metric not enough?

    Good answers cover compounding error, rollout drift, metrics at several horizons and sensitivity to action choices. They may mention uncertainty. Worry if the candidate relies on one reconstruction metric or ignores how errors accumulate.

4 more questions, with what a strong answer covers, are in the Role Kit.

Job description you can post

We are building models that learn how environments change, so that our agents can plan, test and learn without costly real-world trials. We need a researcher who can train world models across text, images, sound and action, and prove they work on unseen trajectories.

What you will do

  • Design and train predictive world models from multimodal and action data
  • Build rollout evaluations on unseen trajectories
  • Measure long-horizon prediction error and diagnose drift
  • Work with planning and agent teams to test downstream usefulness
  • Improve architectures, objectives and data mixes based on evidence
  • Document experiments so others can reproduce them

What you will bring

  • Demonstrated work training generative or predictive models on video, audio or multimodal data
  • Experience with latent dynamics or model-based learning
  • Evidence of careful held-out evaluation, including failure analysis
  • Strong skills in transformer modelling and large-scale training stacks
  • Ability to work with simulators or digital twins
  • Clear written explanation of research results

Nice to have

  • Experience with agent or planning systems that use learned models
  • Published or open work on world models or video prediction
  • Experience building evaluation tooling for generated media

Free to use and adapt. If it saves you time, a link back to this page is appreciated.

Skills that matter

  • Multimodal modellingAdvancedCombine information across modalities
  • Generative mediaAdvancedProduce controllable visual, sound or spatial outputs
  • Transformer modellingWorkingImplement attention-based sequence or multimodal models
  • Simulation and digital twinsWorkingBuild useful virtual representations of environments

Other roles in multimodal and generative world models

Hiring nowMultimodal Research ScientistA Multimodal Research Scientist builds models that connect text, images, sound and action, and designs the benchmarks that show where they break…EmergingGenerative Video ResearcherA Generative Video Researcher builds models that produce video that stays consistent over time and responds to control. They make video generation…EmergingGenerative 3D ResearcherA Generative 3D Researcher builds models and pipelines that produce spatial assets with controllable structure. They matter because generated…

Related roles in other families

Hiring nowPretraining Research EngineerFoundation model pretrainingA pretraining research engineer runs controlled experiments on how large models learn from broad data. The work decides which architectures, data…Hiring nowModel Architecture ResearcherFoundation model pretrainingA Model Architecture Researcher tests changes to how a model is built, to improve its behaviour or efficiency. The work tells a company which…EmergingScaling Science ResearcherFoundation model pretrainingA Scaling Science Researcher works out how model performance changes with compute, data and model size. Their estimates tell a company how to spend…EmergingTraining Dynamics ResearcherFoundation model pretrainingA Training Dynamics Researcher works out why large training runs wobble, stall or diverge, and fixes the cause. Fewer failed runs means less wasted…

How we wrote this

Role definitions are our own editorial work, checked against occupational classifications, standards, research and employer pages. Titles vary by company, and this is a description of the work, not a job advert.

← Multimodal and generative world modelsAll new job families

Start a search

Hiring a World Model Researcher?

Tell us the work. We find the people already doing it, whatever their job title says, and you meet them already interviewed for your role.