NUS, NTU Survey Maps Foundation Models Across Game Lifecycle

A 120-page survey posted to Hugging Face's daily papers on September 16, "AI for Games in the Foundation Model Era," sorts current foundation-model work into six roles across a game's lifecycle: playing and acting, modeling players and games, design, building and maintaining, generating and adapting at runtime, and testing and evaluating. It is authored by researchers at the National University of Singapore and Nanyang Technological University and draws on 439 references.

The authors' central caution reads: "Artifact reuse is not capability transfer." A model that clears one benchmark, they argue, does not automatically clear another; control schemes, rules, engine interfaces, and state representations tend to remain setting-specific, and downstream benefits must be revalidated in target settings.

By the paper's own map, evaluation is well-established for bounded game playing and selected learned environments, and less established for persistent state in learned worlds, repeated software revision, validated player modeling, sustained runtime adaptation, and representative automated testing. That list doubles as a research agenda.

The paper sits at #2 on Hugging Face's daily papers with 38 upvotes, arriving into a run of agent-oriented research our tracker has been logging over the last quarter.