Home / Frontier practices / World Models & Spatial Intelligence
WORLD MODELS / SPATIAL INTELLIGENCE

Models that understand
the world.

World-model teams build representations of environments, motion and interaction that help intelligent systems predict, plan and act. The recruiting challenge sits at the intersection of generative modelling, computer vision, 3D understanding and scalable training.

Impax
01 / THE TECHNICAL LANDSCAPE

Specialism is
the difference.

Search quality depends on understanding what the team is actually building. We map the technical problem first, then identify the skills, research evidence and operating style that matter most for the hire.

Our scope spans the following areas, with each brief shaped around the specific stage, stack and constraints of the business.

  • 01Video generation and prediction
  • 023D scene representations and reconstruction
  • 03Spatial and geometric reasoning
  • 04Diffusion and autoregressive modelling
  • 05Large-scale multimodal pre-training
  • 06Simulation-informed evaluation and planning
02 / REPRESENTATIVE MANDATES

Roles with
real consequence.

These are examples of role types within our specialist coverage, not claims that each role is currently open. For active searches, visit the opportunities board.

  • World Model Research Scientist
  • Generative Video Research Engineer
  • 3D Vision Scientist
  • Foundation Model Training Engineer
  • Multimodal Evaluation Lead
03 / HOW WE ASSESS

Signals over
superlatives.

Our approach starts with the evidence a hiring manager would actually want to see. These are sample questions that help us evaluate depth; they are not universal pass/fail rules.

What has the candidate trained or improved at meaningful scale?+

We separate visual synthesis from spatial reasoning, dynamics and planning. A candidate who has generated convincing video may have a different skill set from one who has built representations useful for action.

Can they reason about the difference between visual plausibility and physically useful prediction?+

Meaningful research signals can include novel architectures, reproducible experiments, training ownership, published artifacts or evidence of improvements under realistic compute constraints.

Do they understand evaluation beyond standard offline benchmarks?+

Evaluation can include temporal consistency, 3D coherence, controllability, simulation fidelity and downstream usefulness. The role’s objective determines which of those matter.

04 / CONNECTED DISCIPLINES
BUILDING IN THIS SPACE?

Let’s discuss the people
who will move it forward.

Speak with Impax ↗
A CLOSER LOOK / ROLE CALIBRATION

The world-model talent map

“World model” can refer to learned representations of space and time, generative simulators, action-conditioned prediction or systems that support downstream planning. The relevant skill profile varies sharply. Hiring for long-horizon video prediction differs from hiring for spatial geometry, multi-view reconstruction or interactive, controllable worlds.

We consider modelling depth, temporal consistency, data curation, evaluation methodology and the intended downstream task. The most informative evidence might be a research paper, a trained system, an interactive demo or rigorous analysis of how a model fails outside curated examples.

REPRESENTATIVE MANDATES

The talent we help identify.

01World-model research scientists
02Spatial intelligence & 3D vision researchers
03Video generation and simulation engineers
04Research engineers for training and evaluation