AI for Games in the Foundation Model Era
Meng Luo, Yanlin Li, Hao Li, Hongzhan Lin, Pengfei Zhou, Tianjie Ju, Ran Zhang, Yeying Jin, et al.
cs.AI
Quick verdict
A six-role survey of AI across the game lifecycle; strong taxonomy and calibrated evidence reading, but no new experiments and a fast-moving preprint base.
Summary
This survey organizes AI-for-games research into six roles defined by the immediate use of AI output: playing and acting; modeling players and games; designing games; building and maintaining games; generating and adapting at runtime; and testing and evaluating games. It reports no new experiments. Instead, three recurring questions structure each chapter: what structure the game or workflow supplies versus what AI learns; which capabilities and artifacts transfer or are reused across settings and roles; and what evidence supports claims at the point of use. Sections 3-8 review specialist and generalist agents, world models and simulators, procedural and language-driven design, engine-integrated development agents, runtime generation, and automated testing, citing figures such as StatePlay's 18.6-point mechanics-fidelity gain over its strongest stateless baseline, Dreamer 4's 29% iron-pickaxe rate in Minecraft, and GameEngineBench's 55.5% best-configuration completion on 110 Unreal tasks. The authors synthesize cross-role connections - trajectories train world models, learned environments train agents, design specifications drive implementations, play and test feedback guides revision - while concluding that control schemes, rules, engine interfaces, state representations, and player contexts remain setting-specific. Evaluation is standardized mainly for bounded game playing; persistent world state, repeated revision, validated player modeling, sustained runtime adaptation, and representative automated testing remain under-evidenced. Limitations: qualitative synthesis only, heavy reliance on recent preprints and corporate announcements, and judgment-dependent role assignments.
Review complete
Assessment at a glance
3
risk flags identified
● high 0
● medium 1
● low 2
5 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Organizing game-AI literature by the immediate use of AI output yields six roles that connect previously separate research threads and clarify what transfers between them.
- Cross-role artifact reuse is demonstrably emerging: trajectories train world models, learned environments supply agent experience, design specifications drive executable implementations, and play or testing feedback guides revision.
- Control schemes, rules, engine interfaces, state representations, and player contexts remain largely setting-specific, so downstream capability claims require validation in the target setting.
- Standardized, execution-grounded evaluation exists mainly for bounded game playing and selected learned environments, while persistent world state, repeated software revision, validated player modeling, sustained runtime adaptation, and representative automated testing remain under-evidenced.
Strengths
Operational, consistently applied taxonomy
Six roles are defined by the immediate use of AI output (Table 1), and §2.1 gives explicit primary/secondary placement rules with worked examples (MarioGPT→Design, Play2Code→Build and Maintain), making the classification reproducible rather than impressionistic.
Disciplined evidence calibration
The Boundary/Transfer/Evidence questions are applied uniformly; §3.1 separates algorithm reuse (DreamerV3's fixed configuration across 150+ tasks) from parameter sharing (Gato, SIMA 2, NitroGen) and held-out-level generalization (Procgen), refusing to collapse them into a single 'generality' claim.
Honest flagging of weak results
Strong triangulation of build-quality evidence (§6.3)
Public maintained index
Weaknesses
No new empirical contribution or meta-analysis
Evidence base skewed to unreviewed material
Judgment-dependent role placement
LLM/VLM-judge metrics cited without independent validation
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This ~80-page survey (arXiv:2609.16679, cs.AI, Sep 2026; NUS/NTU Singapore) organizes AI-for-games research into six roles defined by the immediate use of AI output: playing and acting (§3), modeling players and games (§4), designing games (§5), building and maintaining games (§6), generating and adapting at runtime (§7), and testing and evaluating (§8). The contribution is synthesis, not a new system: Table 1 fixes each role's output, applications, and principal empirical claim, and §2.1 states explicit placement rules (MarioGPT → Design; Play2Code → Build and Maintain). Three questions recur: what structure the game or workflow supplies versus what AI learns; which capabilities and artifacts transfer or are reused; and what evidence supports a claim at its point of use. Four central claims emerge: the six-role lens connects previously separate threads; cross-role artifact exchange is real but partial; control schemes, rules, engine interfaces, state, and player contexts remain setting-specific; and standardized evaluation exists mainly for bounded game playing.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 4 weaknesses, analyzed in detail
- 3 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What are the six roles of AI in games in this survey?
Playing and acting; modeling players and games; designing games; building and maintaining games; generating and adapting at runtime; and testing and evaluating games. Roles are assigned by the immediate use of the AI system's output, with the primary role based on where the system's main empirical claim rests, and secondary roles recorded via cross-references.
Does this paper present new AI systems or experiments?
No. It is a survey: it organizes and critically reviews existing systems, benchmarks, and evaluation evidence, and synthesizes cross-role connections such as using gameplay trajectories to train world models. Its contributions are the taxonomy, the transfer/reuse analysis, and the evidence assessment - not new measurements.
What is the paper's main conclusion about evaluation in game AI?
Evaluation is most standardized and execution-grounded for bounded game playing and some learned environments. Persistent world state, repeated software revision, validated player modeling, sustained runtime adaptation, and representative automated testing still lack established evidence, and because control schemes, rules, engine interfaces, and state representations are largely setting-specific, results often do not transfer without re-validation in the target game.
Is there a code or resource repository for this survey?
Yes. The authors maintain a project page and a GitHub repository (awesome-ai-for-games) containing a system index that records primary and secondary roles for surveyed systems and named components, plus per-system comparison tables referenced in the paper's appendix.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Extracted text truncates around the start of Section 7; Sections 8-11 (testing, benchmark synthesis, cross-role discussion, challenges) were assessed only from the table of contents, abstract, and introduction.
- Figures are available only as captions; the timeline (Fig. 3), role diagram (Fig. 2), and in-game screenshots could not be verified visually.
- The reference list was not included in the extraction; citation balance and recency of later chapters could not be audited.
- The metadata abstract cuts off mid-sentence; closing abstract claims were reconstructed from the in-paper abstract.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.