AI for Games in the Foundation Model Era

AI for Games in the Foundation Model Era. By Meng Luo and nine others, from the National University of Singapore and Nanyang Technological University. Published on the fifteenth of September. Its code repository is public and real: thirteen source files.

Foundation models, together with advances in learned game-world models, are reshaping AI across the whole lifecycle of a game. Systems now do much more than play games. They model players and game dynamics, support design and development, adapt the experiences players see at runtime, and evaluate the artifacts that result. These directions have mostly evolved separately. That separation has obscured which capabilities transfer across settings, and which stay tied to particular games, engines, interfaces or player populations.

The survey sorts the literature into six roles, defined by the immediate use of the AI's output. The first is playing and acting. The second, modeling players and games. The third, designing games. The fourth, building and maintaining games. The fifth, generating and adapting at runtime. And the sixth, testing and evaluating games. The roles follow a game roughly from conception through to release and play.

For each role, the same four questions are asked, which is what makes the roles comparable. What structure does the game or workflow supply? What does the AI learn or produce? Which capabilities and artifacts transfer across settings and roles? And what evidence supports the claims being made? The fourth question is the one the rest of the survey keeps returning to, because a capability that works in one game is not yet evidence that it works in another.

Several connections run between the roles. Trajectories from play train world models. Learned environments give agents experience. Design specifications drive executable implementations. And feedback from play or testing guides revision. In each case the output of one role becomes the input of another. Read together, these connections suggest the six roles are less separate than the literature's history makes them look, which is the premise the survey set out from.

Against that, much does not travel. Control schemes, rules, engine interfaces, state representations and player contexts often remain specific to one setting. So a downstream claim has to be validated in the target setting where it will be used, not assumed from the setting where it was shown.

Evaluation is uneven across the roles. It is most standardized for bounded game playing, and for selected learned environments. It is less established for persistent state in learned worlds, for repeated software revision, for validated player modeling, for sustained runtime adaptation, and for representative automated testing.

The authors name one central challenge that ties the survey together. Outputs and capabilities should be reused or transferred across roles, and each time they are, the evidence that they work has to be established again, in the game-specific context where they end up.