Foundation Models for Game World and Player Modeling

What are world models for games?

World models for games learn representations of game dynamics, state transitions, environments, and sometimes player behavior. A model may predict the next state for planning, provide a learned simulator for policy training, compare possible futures, or generate an interactive experience directly from actions and prior observations.

Different world models preserve different information. Video-based models may prioritize plausible frames, while structured or latent models may emphasize mechanics, controllable state, or decision-relevant features. Visual realism alone does not prove that collisions, inventory, health, rules, or long-term consequences are correct, and a short coherent clip does not establish persistent state across revisits.

Player models address a connected but separate question: forecasting choices, preferences, skill, affect, or trajectories from behavioral data. Those predictions can support analytics and adaptation, but do not decide what a game should change. Evaluation should match the intended consumer—planning, training, simulation, generation, or player analysis—and test perceptual quality, mechanics fidelity, state persistence, uncertainty, policy transfer, and performance on held-out players or games.

This topic is surveyed in AI for Games in the Foundation Model Era.