AIAny
Icon for item

AI for Games in the Foundation Model Era

Surveys how foundation models reshape AI across the game lifecycle, organizing literature into six roles (play & act; model players & games; design; build & maintain; runtime generation & adaptation; test & evaluate). Emphasizes cross-role links, limits to transfer, and gaps in evaluation and persistent state.

Introduction

Foundation models and learned world models are widening what AI can do in games — not just play, but model players and worlds, drive design, write and repair code, generate runtime content, and automate testing. The paper argues this shift creates new opportunities for reuse across roles but also exposes how engine, interface, rule, and player-context specificity block straightforward transfer of capabilities.

Key Findings
  • Role-based taxonomy: Organizes the literature into six operational roles, clarifying what AI output is used for and how workflows constrain learning and reuse. This makes it easier to compare systems by immediate function rather than model family.
  • Cross-role connections and limits: Play trajectories train world models and learned environments seed agent experience, while design specs can drive executable implementations — yet many artifacts fail to transfer without revalidation because of control schemes, state representations, and engine interfaces.
  • Evidence gaps: Evaluation is relatively standardized for bounded game playing, but persistent learned worlds, validated player modeling, sustained runtime adaptation, and representative automated testing remain underdeveloped.
  • Practical concerns: Raises caution that artifact reuse is not the same as capability transfer; downstream benefits require new evidence in target settings.
Who it's for and tradeoffs

Great fit if you need a structured map of recent AI-for-games work and a critical lens on how foundation models intersect with production workflows. Useful for researchers building generalist agents, designers exploring AI-native mechanics, and engineers integrating generative models into runtime systems. Look elsewhere if you want implementation recipes, step-by-step engineering guidance, or benchmarking code — the paper is a conceptual survey and synthesis rather than an engineering tutorial.

Information

  • Websitearxiv.org
  • AuthorsMeng Luo, Yanlin Li, Hao Li, Hongzhan Lin, Pengfei Zhou, Tianjie Ju, Ran Zhang, Yeying Jin, Mong-Li Lee, Wynne Hsu
  • Published date2026/09/15

More Items

Systematically studies how language and vision interact during unified multimodal pretraining, identifies mechanisms that enable modality synergy versus competition, demonstrates the benefit of early joint training, and derives efficient pretraining recipes validated at scale.

Presents Metis, a prototype memory foundation model that embeds a persistent native memory state into the backbone so historical experience is compressed and accessed via memory attention. Key features: forward-only, gradient-free online memory updates; memory-specific mid-training objectives; and a dual text/code memory design.

Empirically studies how transformer-based native multimodal pre-training scales under fixed compute, deriving compute- and data-allocation power laws and an efficiency frontier for model size, token count, and data mixture; evaluates cross-modal transfer and multimodal in-context learning.