Research Scientist (Physical AI/Interpretability)

About World Mechanics World Mechanics is a commercial R&D neo-lab developing interpretable foundation models of and for the physical world. We study what mechanisms world models learn about the systems they are trained on, and how to make those mechanisms more faithful to the underlying causal structure of those systems. We believe that models which learn interpretable, causal abstractions can provide a more principled foundation for explaining predictions, verifying model behavior, and monitoring and controlling models in deployment. Our goal is to build general-purpose world models that serve as reliable simulators of physical systems and can be deployed broadly across applications ranging from prediction and maintenance to robotics and scientific discovery. We focus on three connected research directions:

• white-box evaluation of physical models,

• training intrinsically interpretable foundation models for the physical world,

• building interpretability-based simulators for physical systems.

What you’ll do

• Build interpretability tooling and experiments to find, test, and validate internal representations in physical AI models (e.g., motion, dynamics, object permanence, latent system state).

• Design “white-box evaluation” methods: detect internal failure signatures, trace model errors to representations/computations, and create actionable diagnostics.

• Contribute to interpretability-informed training loops (train → discover structure → evaluate → incorporate → repeat), turning interpretability discoveries into training signals or inductive biases.

• Work across video, sensor, time-series, and other multimodal data, and help determine what useful representations look like across physical and industrial systems.

• Collaborate with the team (and with customers!) to ensure research stays grounded in real deployment constraints and commercial value.

• Work toward publications at top conferences and journals, and publish open-source artifacts from our research agenda when aligned with our commercial and customer goals.

What we’re looking for

• Strong ML research ability with demonstrated execution (papers, strong open-source, or substantial internal research impact).

• Fluency in deep learning fundamentals and the ability to prototype quickly and iterate empirically.

• Comfort working in ambiguity: you can pick good bets, run clean experiments, and update quickly.

• Strong communication and epistemic humility; you like productive debate without ideology or politics.

• Interest in commercialization and product pull; you care whether the research becomes a real capability.

Nice to have

• Mechanistic interpretability experience (circuits, feature discovery, attribution/causal methods, eval design).

• Work on multimodal models, video models, robotics/embodied AI, or physical time-series.

• Experience building research infrastructure (training, eval harnesses, data tooling).

Our values

Empirical Rigor and Intellectual Openness We approach research with an open mind and let evidence guide our decisions. We test assumptions, update our views when the facts change, and focus our efforts where they can have the greatest impact. We value clear thinking and practical progress over allegiance to particular schools of thought or debates that are not supported by meaningful evidence. Ownership and Initiative We seek people who are energized by the pace, autonomy, and ambiguity of an early-stage company. Our team members take responsibility beyond narrow job descriptions, move quickly from ideas to execution, and reliably carry work through to completion. We value genuine commitment to the mission and the desire to build, not simply to participate. Seriousness, Kindness, and Mutual Respect We aim to build a team of thoughtful, optimistic, and grounded people who care deeply about both the quality of their work and how they treat others. We keep ideological and political debates—including factional debates around AI safety and capabilities—outside the workplace and maintain our focus on the work itself. We welcome substantive disagreement while holding a high bar for respect, integrity, and professional conduct. Research with Real-World Impact We pursue research with the intention of translating it into useful products and capabilities. We value researchers who are enthusiastic about commercialization and want to see their work succeed beyond the research environment. Staying close to product, engineering, and—when useful—customers and real-world settings helps ensure that our research remains grounded in genuine needs and constraints.