Boundary
What is supplied by the game or workflow, and what is assigned to AI?
THE FOUNDATION MODEL ERA
AI for Gamesin the Foundation
Model Era
The next move is
more than playing.
A visual companion to the survey of how foundation models play, model, design, build, adapt, and test games.
Final paper · 120 pages · Concepts, evidence, and an open reading collection
01 / A WORLD IN THE MAKINGConcept artwork
444references
06roles across the lifecycle
03questions across every role
VIDEO INTRODUCTION
The paper,
in 90 seconds.
A concise visual introduction to the six research roles, their connections, and the evidence needed to assess progress across the game lifecycle.
01:30 · Full presentation · Sound available01 THE BIG PICTURE
The game is no longer
only given.
Game AI has never been limited to playing. The foundation-model era makes the field’s wider shape impossible to miss.
AI can now act as a player or character, model a world or its players, propose content and rules, operate inside a game engine, change a live experience, and produce testing evidence. The same pretrained model family may touch several of these tasks, but each role still depends on different interfaces, constraints, and forms of evidence.
The survey therefore follows the immediate use of AI output. This separates a proposed mechanic from its implementation, a player forecast from the adaptation it informs, and a tester’s action from the verdict it supports.
See the organizing map02 THE ORGANIZING MAP
Six roles around
a playable game.
The roles classify what an output is used for. The connectors show selective inputs and feedback—not a required architecture.
Six uses of AI output around an executable, player-facing game.
THREE QUESTIONS, REPEATED ACROSS EVERY ROLE
These questions turn the foundation-model era from a date label into a common analytical lens.
Transfer & reuse
Which capabilities transfer, which artifacts can be reused, and what remains setting-specific?
Evidence
What claims does evaluation support where the output is used?
FROM THE MANUSCRIPT
The timeline traces selected lineages; the two-part knowledge map locates the technical directions and cited systems discussed in the survey.
03 THE ROLE ATLAS
Inside each
research role.
Each chapter follows a different output, operating context, and empirical claim. Select a role to see its research structure and synthesis.
01 / PLAY AND ACT
AI that Plays and Acts
AI selects actions, plans, or messages inside a game. Foundation models broaden perception and planning, while the path to native controls remains game-specific.
- Output
- Actions · plans · messages
- Applications
- Players · teammates · NPCs
- Main claim
- Action quality
CHAPTER STRUCTURE
- Player and generalist agents
- Learning and control hierarchies
- Test-time adaptation and memory
- Opponents, teammates, NPCs, and companions
SURVEY SYNTHESIS
- A common interface relocates the learning problem.
- Adaptation includes the information supplied.
- Playing well with others is a separate transfer problem.
04 CROSS-ROLE DISCUSSION
What connects—and
what actually transfers.
An implemented exchange shows that information can move between roles. It does not by itself show that competence transfers or that the receiving workflow improves.
Broader interfaces, game-specific structure
Language, multimodal, and code interfaces connect high-level intent to constrained plans, tools, actions, or learned dynamics. Reliability still depends on what the receiving game or component must execute.
Artifact reuse is not capability transfer
A trajectory, specification, profile, or test trace can be consumed downstream without proving competence under a new game, engine, interface, player population, or task. Transfer requires naming what changed and testing whether the capability survives.
Downstream benefit needs its own test
Compatibility, actual use, and benefit are separate checks. A connection earns a stronger claim only when the intended downstream outcome improves against a suitable baseline, with cost and failure propagation made visible.
Interaction tracesLearned environmentsSpecifications & rulesExecution & test evidencePlayer models & profiles
AN EMERGING CONNECTION · DESIGN / BUILD / MODEL / RUNTIME
One route separates executable game state from generated appearance: a coding agent builds a graybox, the engine executes rules and updates state, and a state-conditioned model renders what players see. Programmable World Model ↗ demonstrates a related loop with a lightweight engine; AlayaRenderer-Flash ↗ brings neural rendering to a live game engine. The full UE/Unity development-to-rendering workflow is a research direction, not an already validated end-to-end system.
05 EVIDENCE & RESEARCH DIRECTIONS
The claim decides
the test.
The six roles do not share a single measure of success. Each card pairs the main evaluation target with the boundary that current research still needs to cross.
01
Play & Act
Measure completion, progress, efficiency, and coordination.
Next test vary appearance, controls, rules, and partner conventions under fixed interaction and real-time budgets.
02
Model Players & Games
Measure visual, mechanics, and persistent-state fidelity; reference-game policy return; and held-out player behavior.
Next test revisit changed worlds and distinguish stable player tendencies from temporary context.
03
Design
Measure validity, diversity, constraints, and designer control.
Next test follow real revision work, retained alternatives, correction effort, and the resulting player experience.
04
Build & Maintain
Measure launchability, requirements, repair, and regression.
Next test use multi-version projects and evaluate whether a new developer can reproduce, revise, and hand off the work.
05
Generate & Adapt at Runtime
Measure latency, consistency, controllability, and player outcomes.
Next test run repeated sessions with non-adaptive controls, resumed saves, model updates, and reversible interventions.
06
Test & Evaluate
Measure validated defects, coverage, verdict accuracy, and reproduction cost.
Next test combine diverse exploration with independent oracles, held-out defect families, and relevant player samples.
Three tests, not one score.
Perceptual quality asks whether generated video looks and flows plausibly. Mechanics correctness checks whether actions produce rule-consistent changes. Persistent state asks whether those facts survive revisits and long interaction. FVD measures video-feature distributions; when reference game or engine state is available, health, collisions, inventory, and event traces can be checked separately.
A new map is not a new game.
Changing a map, a mode, or a game tests different things. A new scene may probe visual transfer; an altered control scheme probes action transfer; unfamiliar rules require reasoning that neither demonstrates. Cross-game claims should name the held-out unit and report the behavior that survived it.
PLAY NINE AI-CRAFTED WORLDS
Don’t just read it.
Play it.
Nine browser games span physics puzzles, strategy, racing, platforming, and compact experimental worlds. Start with three new multi-stage adventures, then explore six earlier prototypes.
WORLD 01
Physics puzzle6 fortresses
Cloud Sling 云岛弹射
Launch three companions with distinct mid-air abilities, expose structural weak points, and topple six increasingly elaborate sky fortresses.
DesignBuildPlay
Play now 
WORLD 02
Lane defense18 waves
Sprout Guard 芽芽保卫战
Grow energy, combine six plant roles, upgrade defenses, and hold five lanes against increasingly coordinated mechanical swarms.
DesignBuildPlay
Play now 
WORLD 03
3D arcade racing3 tracks
Neon Kart 霓虹卡丁
Race six karts across three circuits, charge nitro through drifts, collect items, and turn every corner into an overtaking opportunity.
DesignBuildPlay
Play now 
WORLD 04
Spatial puzzle3 levels
Moonlight Garden 月光花园
Push planters onto moonlit goals, hold pressure plates, open the bridge, and undo mistakes while solving three compact spatial puzzles.
DesignBuildTest
Play now 
WORLD 05
Turn-based strategy3 levels
Aurora Outpost 极光哨站
Read the next storm footprint, manage scarce energy, activate every beacon, and return safely using dash, shield, and wait.
DesignBuildTest
Play now 
WORLD 06
Platform adventureAI-created
Lulu's Snowy Letter 露露的雪夜来信
Guide a winged dragon across moonlit platforms, collect starlight, and use two model-composed abilities to reach the portal.
DesignBuildPlay
Play now 
WORLD 07
Ocean rescueAI-revised
The Lighthouse Is Still Shining 灯塔还亮着
Search a snowy archipelago, bring stranded companions aboard, and return them safely to the lighthouse in limited-capacity trips.
DesignBuildPlay
Play now 
WORLD 08
Survival arenaAI-revised
Garden Watch 花园守望
Dodge persistent pursuers, gather starlight, and combine a repelling pulse with a short shield to protect the garden.
DesignBuildPlay
Play now 
WORLD 09
Platform journeyAI-revised
A Letter from the Clouds 云端来信
Traverse a revised route of floating islands, avoid crimson traps, collect starlight, and reach the final portal.
DesignBuildPlay
Play now Playable in the browserNo installation
06 PROJECTS, IN VIEW
Ideas made
concrete.
Six roles, seen through public project imagery. Each example links to its original source and is described at the level its evidence supports.
Learning to act.
Learning to cooperate.
From specialist policies to language-guided agents, this literature studies how AI perceives a game, chooses actions, and coordinates with other players.
Generalist policiesPlanning & memoryHuman–AI teamwork
IN THE IMAGE
Explore this role GT Sophy brings learned racing behavior into Gran Turismo 7, connecting research in real-time control with a game people can play.
Illustrations from the cited projects; rights remain with their respective owners.
← → to navigate collections07 THE READING ROOM
Follow your curiosity.
Search 436 current manuscript references and eight separately tracked public additions by topic, year, or keyword. Role filters distinguish primary research roles from supporting foundations and context.
444 references in the bibliography
Download BibTeX ↓Newest first
No matching papers.
Try another title or topic, or reset your filters.
08 AN OPEN COLLECTION
The next discovery
might be yours.
Know a paper that belongs here? Help make the collection more useful to researchers, designers, and developers.