Applied Research Scientist, World Model

Zillow

4d•$185,800 - $312,400•Remote

About The Position

Zillow’s Agentic Foundations team is a group of 15–20 applied scientists within its broader Agentic AI organization. The team is building foundational AI technologies for future home-shopping experiences, including the Zillow World Model: an uncertainty-aware model connecting housing supply, demand, and macroeconomic conditions. The model is intended to provide consistent projections for multi-step consumer questions and serve those answers to Voyager, Zillow’s AI Mode. The role builds the World Model’s demand layer by transforming large-scale shopper behavior into structured demand representations, including attention-to-demand conversion. It involves developing models for latent market states and dynamics, estimating willingness to pay and market-clearing behavior, and applying causal methods to establish when counterfactual claims are justified. The scientist will contribute to model evaluation and independently lead open-ended research from problem definition through production impact, while providing technical leadership and mentorship. The role partners with the World Model working group and product, engineering, platform, and data teams. This position is categorized as remote. Remote employees have no permanent corporate office workplace and work from a physical location of their choice that must be identified to the Company. U.S. employees may live in any of the 50 states, with limited exceptions. The position is eligible for equity awards. Listed annual base salary ranges vary by location: $195,600–$312,400 in California, Connecticut, Maryland, Massachusetts, New Jersey, New York, Washington state, and Washington, DC; and $185,800–$296,800 in Colorado, Hawaii, Illinois, Maine, Minnesota, Nevada, Ohio, Rhode Island, Vermont, and Virginia. Actual compensation depends on experience, performance, and location.

Requirements

PhD in Computer Science, Machine Learning, Statistics, Economics, or a related quantitative field.
Hands-on experience building latent state-space or world models, such as RSSM/Dreamer-family models, JEPA-style latent prediction, or event-sequence transformers, including belief-state representations, stochastic transition models with calibrated uncertainty, and multi-step rollout evaluation.
Demonstrated ability to turn large-scale user behavior data into structured demand representations, such as utility or willingness-to-pay estimation and funnel or conversion modeling, while correcting for selection bias in platform-observed populations.
Practical causal inference experience in observational settings, including quasi-experiments, instruments, or policy discontinuities, and judgment about whether learned correlations support counterfactual claims.
A proven record of delivering open-ended research projects, scoping ambiguous problems, and providing technical leadership to scientists and MLEs.
Strong programming skills, especially in Python, and experience with ML frameworks such as PyTorch or TensorFlow.

Nice To Haves

Economics fluency, especially in market microstructure, search-and-matching, structural or discrete-choice demand models, and equilibrium feedback.
Experience with agent-based or market simulation, or simulation-based reinforcement learning.
Published work in world models, model-based reinforcement learning, causal inference, or structural demand estimation.
Experience with multimodal representation learning or bitemporal, leakage-safe training data.
Familiarity with two-sided marketplaces, real estate, or long-horizon consumer journeys.

Responsibilities

Build the World Model’s demand layer by turning large-scale shopper behavior, including search, engagement, and session trajectories, into structured demand representations.
Model explicit attention-to-demand conversion.
Design and train latent state-space, multi-tower, and/or sequential dynamics models with calibrated uncertainty for multi-step rollouts and forward and backward planning.
Estimate willingness to pay, per-home demand curves, and market-clearing dynamics, connecting learned representations to economic quantities.
Apply causal identification to observational data and define when the model must abstain from claims unsupported by causal evidence.
Contribute to the World Model evaluation battery, including state probes, calibrated multi-target forecasts, rollout drift, causal checks, and out-of-regime replays.
Drive ambiguous research independently from end to end.
Provide technical leadership and mentorship in translating research into production impact.