AI for Games in the Foundation Model Era

Meng Luo; Yanlin Li; Hao Li; Hongzhan Lin; Pengfei Zhou; Tianjie Ju; Ran Zhang; Yeying Jin

cs.AISep 15, 2026

AI for Games in the Foundation Model Era

Authors: Meng Luo, Yanlin Li, Hao Li, Hongzhan Lin, Pengfei Zhou, Tianjie Ju, Ran Zhang, Yeying Jin,

Organizations: National University of Singapore · Nanyang Technological University

Abstract

Foundation models, alongside advances in learned game-world models, are reshaping AI across the game lifecycle. Beyond playing games, recent systems model players and game dynamics, support design and development, adapt player-facing experiences at runtime, and evaluate resulting artifacts. Yet these directions have evolved largely separately, obscuring which capabilities transfer across settings and which remain tied to particular games, engines, interfaces, or player populations. We organize the literature into six roles according to 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. For each role, we examine what structure is supplied by the game or workflow, what AI learns or produces, which capabilities and artifacts transfer across settings and roles, and what evidence supports the claims. We identify cross-role connections: trajectories train world models, learned environments provide experience for agents, design specifications drive executable implementations, and play or testing feedback guides revision. However, control schemes, rules, engine interfaces, state representations, and player contexts often remain setting-specific, so downstream claims require validation in the target setting. Evaluation is most standardized for bounded game playing and selected learned environments, while persistent state in learned worlds, repeated software revision, validated player modeling, sustained runtime adaptation, and representative automated testing remain less established. The central challenge is to reuse or transfer outputs and capabilities across roles while re-establishing evidence for effectiveness in the game-specific contexts where they are used.

Figures & tables

Figure 2: Six uses of AI output around an executable, player-facing game. The connectors indicate selective inputs and feedback rather than a mandatory pipeline.

Role	Output	Applications	Claim
Play and Act	actions; plans; messages	players; teammates; NPCs	action quality
Model Players and Games	states; transitions; player forecasts	planning; simulation; player models	predictive quality ; policy transfer
Design	levels; rules; story structures	PCG; automated design; co-creation	validity; control
Build and Maintain	code; scenes; project edits	engine agents; debugging; repair	working software
Generate and Adapt at Runtime	dialogue; quests; live content	characters; adaptive narrative	consistent content ; player response
Test and Evaluate	traces; verdicts; diagnoses	playtesting; verification; judging	state coverage ; verdict accuracy

Table 1: Six roles: outputs, applications, and the principal empirical claim.

Figure 3: Timeline of representative AI-for-games systems across the six roles. Selected systems, benchmarks, and methods reviewed in this survey are arranged by year and colored according to their primary role in our taxonomy.

Figure 6: Research directions for AI that plays and acts: specialist-to-generalist policies, test-time adaptation, control hierarchies, and NPCs and teammates.

Panel of Figure 7 — Figure 7: Specialist agents built for a single title (a–c) and policies whose parameters transfer across commercial games (d–f): (a) AlphaStar, StarCraft II ( Vinyals et al., 2019 ) ; (b) OpenAI Five, Dota 2 ( Berner et al., 2019 ) ; (c) Honor of Kings AI ( Ye et al., 2020 ) ; (d) NitroGen, The Witcher 3 ( Magne et al., 2026 ) ; (e) NitroGen, Rocket League ( Magne et al., 2026 ) ; (f) SIMA 2, No Man’s Sky ( SIMA Team et al., 2025 ) .

Setting	Reused	Game-specific work	Examples
Per-game training	algorithm; settings	policy training; dynamics learning	DQN; AlphaZero; DreamerV3 [ 222 , 290 , 124 ]
Formal specification	solver; formal interface	rules; forward model	GGP; GVGAI [ 107 , 249 ]
Parameter sharing	policy weights	held-out testing; fine-tuning	Gato; Multi-Game DT; SIMA 2; NitroGen [ 263 , 171 , 292 , 206 ]
Test-time adaptation	base model; retrieval	demonstrations; rules; exploration	REGENT; Code World Models; Twin [ 300 , 172 , 295 ]

Table 2: Recurring training and evaluation settings for game-agent generality.

Panel of Figure 8 — Figure 8: An intention passing down the control hierarchy, from a stored skill (a) to a planner-selected target executed by the controller (b): (a) Voyager, Minecraft ( Wang et al., 2023a ) ; (b) ROCKET-1, Minecraft ( Cai et al., 2025b ) .

Panel of Figure 9 — Figure 9: Test-time adaptation by retrieving demonstrations for an unseen game (a) and by carrying episodic state through an episode (b): (a) REGENT, Ms. Pac-Man ( Sridhar et al., 2025 ) ; (b) EMemBench, Crafter ( Li et al., 2026b ) .

Panel of Figure 10 — Figure 10: Task-dependency-based multi-agent coordination (a) and social agents whose behaviour emerges from interaction (b–d): (a) VillagerAgent, Minecraft ( Dong et al., 2024 ) ; (b) Generative Agents, Smallville ( Park et al., 2023 ) ; (c) Project Sid (PIANO), Minecraft ( Altera. AL et al., 2024 ) ; (d) Comme il Faut, Prom Week ( McCoy et al., 2013 ) .

Figure 11: Research directions for modeling players and games: planning, training, and simulation; world-state representation; player behavior modeling; and long-horizon interaction.

Panel of Figure 12 — Figure 12: Learned simulators that a player interacts with directly; each frame is produced in response to the input shown: (a) GameNGen, DOOM ( Valevski et al., 2025 ) ; (b) Genie 2, generated street ( Parker-Holder et al., 2024 ) ; (c) Genie 3, generated world ( Parker-Holder and Fruchter, 2025 ) ; (d) Hunyuan-GameCraft-2, generated voxel world ( Tang et al., 2025a ) .

Panel of Figure 13 — Figure 13: Persistence in learned simulators: scenes that remain consistent when revisited (a–c) and game state predicted as explicit variables (d–f): (a) ReWorld, castle interior ( Chen et al., 2026d ) ; (b) Context-as-Memory, Japanese townscape ( Yu et al., 2025a ) ; (c) WorldCam, FPS map ( Nam et al., 2026 ) ; (d) StatePlay, Street Fighter III ( Lin et al., 2026c ) ; (e) Marionette, monster hunt ( Meng et al., 2026 ) ; (f) WorldMind, Hollow Knight ( Deng et al., 2026 ) .

Panel of Figure 14 — Figure 14: Games in which player experience (a), retention (b), and simulated players (c) are modelled: (a) Pinitas et al., The Division 2 ( Pinitas et al., 2023 ) ; (b) Roohi et al., Angry Birds Dream Blast ( Roohi et al., 2020 ) ; (c) Generative personas, Solid Rally ( Barthet et al., 2022 ) .

Panel of Figure 15 — Figure 15: Action interfaces of generative simulators: device-action conditioning (a–b), natural-language entity commands (c–d), and player actions with NPC strategy conditioning (e–f): (a) SCOPE, open-world FPS ( Tong et al., 2026 ) ; (b) SCOPE, temple-gate scene ( Tong et al., 2026 ) ; (c) Incantation, Elden Ring ( Zhu et al., 2026a ) ; (d) Incantation, Elden Ring ( Zhu et al., 2026a ) ; (e) ReactiveGWM, Street Fighter II ( Wang et al., 2026a ) ; (f) ReactiveGWM, Street Fighter III ( Wang et al., 2026a ) .

Figure 16: Research directions for AI that designs games: assets, levels, and worlds; rules and mechanics; and narrative and co-creative design.

Panel of Figure 17 — Figure 17: Generated content that must remain playable: search-based (a), language-conditioned (b), and engine-compatible 3D generation (c): (a) MarioGAN, Super Mario Bros. ( Volz et al., 2018 ) ; (b) MarioGPT, Super Mario Bros. ( Sudhakaran et al., 2023 ) ; (c) Xu et al., Left 4 Dead 2 ( Xu and Verbrugge, 2025 ) .

System	Generated object	Feedback / checks	Supplied structure
MarioGPT [ 302 ]	tile sequence	A*; prompt features	tiles; movement rules
DreamCraft [ 88 ]	block-aligned 3D scene	text alignment; adjacency losses	Minecraft block vocabulary
CubePart [ 416 ]	part-based 3D asset	part control; geometry	schemas; external behavior
GAVEL [ 326 ]	Ludii rule program	compile; simulated play; diversity search	game-description language
ScriptDoctor [ 89 ]	rules; tile levels	compiler; BFS; iterative repair	DSL; engine; solver budget
KNUDGE [ 352 ]	branching dialogue	lore and quest consistency	lore; quest goals

Table 3: Design methods compared by output, feedback, and supplied game structure.

Panel of Figure 18 — Figure 18: Narrative generated under authored constraints (a) and a designer-edited level evaluated through persona-based playtesting (b): (a) NarrativeGenie, Unity episode ( Kumaran et al., 2024 ) ; (b) a later MiniDungeons extension of Sentient Sketchbook ( Liapis et al., 2013 ; Liapis et al., 2015 ) .

Figure 19: Research directions for AI that builds and maintains games: code, scenes, and engine projects; tool-using development agents; execution, debugging, and repair; and revision, maintenance, and handoff.

Panel of Figure 20 — Figure 20: Complete Godot projects constructed by coding agents and verified by running them: (a) GameCraft-Bench, Cozy Harbor Delivery ( Luo et al., 2026 ) ; (b) GameCraft-Bench, horror game ( Luo et al., 2026 ) .

Panel of Figure 21 — Figure 21: Games produced by development agents that observe the running build (a) or specialise a coding model (b): (a) Play2Code, Garden Guard ( Huang et al., 2026a ) ; (b) OpenGame, generated brawler ( Jiang et al., 2026b ) .

Figure 22: Research directions for AI that generates and adapts at runtime: generative content for dialogues, props, and tasks; long-term memory across characters, state agreement, and shared sessions; dynamic rules and mechanics with engine-side validation; and personalization through affect models, difficulty adaptation, and adaptive characters.

Panel of Figure 23 — Figure 23: Runtime generation in two settings: a question-answering exchange with a code-enabled NPC (a) and generated narrative content integrated into game progression (b): (a) Craft an Iron Sword, Minecraft ( Volum et al., 2022 ) ; (b) PANGeA, Dark Shadows ( Buongiorno et al., 2024 ) .

Panel of Figure 24 — Figure 24: Rules and objects generated during play and executed by the engine: a domain-specific rule language (a–b) and schema-constrained objects (c): (a–b) Real-Time World Crafting, Latent Space ( Drake and Dong, 2025 ) ; (c) Roblox Cube ( Singh, 2026a ) .

Panel of Figure 25 — Figure 25: Runtime personalization and companion systems: personalised narrative (a), adaptive characters (b), an AI teammate in a limited-time public beta (c), and AI squadmates in a closed-playtest prototype (d): (a) Interactive drama (Wu et al.), waiting-room scene ( Wu et al., 2025 ) ; (b) Figueiredo et al., detective prototype ( Figueiredo and Elumeze, 2025 ) ; (c) PUBG Ally ( KRAFTON, 2026 ) ; (d) Ubisoft Teammates, first-person prototype ( O’Connor, 2025 ) .

Figure 26: Research directions for AI that tests and evaluates games: automated playtesting through exploration objectives, procedural personas, and pixel- or GUI-based agents; mechanic verification with executable oracles and coverage-guided regression selection; human-likeness and player coverage; and model-based judgement with QA benchmarks, agreement metrics, and meta-evaluation.

Panel of Figure 27 — Figure 27: Automated playtesting, from persona-driven exploration (a) to learned oracles for visual defects (b–c), screen-only testers (d), and production deployments (e–f): (a) Developing personas, DOOM ( Ariyurek et al., 2023 ) ; (b) Ling et al., glitch sample ( Ling et al., 2020 ) ; (c) Macklon et al., HTML5 canvas game ( Macklon et al., 2022 ) ; (d) Inspector, FPS test arena ( Liu et al., 2022 ) ; (e) EA SEED AutoPlayer, Dead Space zero-gravity navigation ( Gillberg et al., 2023 ) ; (f) iv4XR, Space Engineers ( Prasetya et al., 2022 ) .

Panel of Figure 28 — Figure 28: Benchmark cases in which a model-based judge must decide whether a screenshot contains a defect: (a) VideoGameQA-Bench, co-op FPS heist ( Taesiri et al., 2025 ) ; (b) VideoGameQA-Bench, Rise of the Tomb Raider ( Taesiri et al., 2025 ) ; (c) GlitchBench, The Witcher 3 ( Taesiri et al., 2023 ) ; (d) GlitchBench, Gotham Knights ( Taesiri et al., 2023 ) .

Role	Targets and example measures	Interpretation boundary
Play and Act	Completion; progress; action efficiency	Protocol-specific; transfer across rules, controls, or timing remains untested
Model Players and Games	World-state fidelity; reference-game policy return; player prediction or trait recovery	Short-horizon fidelity does not establish persistence; synthetic traits do not establish real-player validity
Design	Validity; diversity; constraint satisfaction; designer control	Does not establish authoring efficiency or intended player experience
Build and Maintain	Launchability; requirement satisfaction; repair; regression	Does not establish long-term maintenance or developer handoff
Generate and Adapt at Runtime	Latency; consistency; controllability; player outcomes	Single-session evidence; adaptation effects require suitable controls
Test and Evaluate	Coverage; validated defects; verdict accuracy; reproduction cost	Coverage is not correctness; sampled behavior is not population validity

Table 4: Evaluation targets and interpretation boundaries across the six roles.

Model / player	Progress (%) $\uparrow$	Success (%) $\uparrow$
Computer use: keyboard and mouse commands
Seed-1.8	39.8	20.0
Claude Sonnet 4.6	38.3	19.4
Gemini 2.5 Computer Use	36.1	16.5
OpenAI Computer Use	35.8	16.5
Qwen3-VL-Plus	33.6	15.9

Table 5: GameWorld results for 18 model–interface configurations and two human reference participants across 170 tasks in 34 games ( Ouyang et al., 2026 ) .

(a) Standard harness
Model	Reasoning effort	Score (%) $\uparrow$
GPT-6 Astra	Max	62.71
Claude Opus 5	High	30.16
GPT-5.6 Sol	Max	7.78
Grok 4.6	XHigh	2.11
GPT-5.6 Terra	Max	0.80

Table 6: ARC-AGI-3 results on the Semi-Private split ( ARC Prize Foundation, 2026b ; ARC Prize Foundation, 2026a ; ARC Prize Foundation, 2026f ; ARC Prize Foundation, 2026c ; ARC Prize Foundation, 2026e ; ARC Prize Foundation, 2026d ) . The first panel reports the highest Standard-harness score for each listed model at the reasoning-effort setting shown; the second compares harnesses for GPT-6 Astra.

Target	What to check	Suitable evidence
Perceptual quality	Appearance; motion continuity; visual diversity	FVD; perceptual scores; human ratings
Mechanics correctness	Legal transitions; health/resource changes; contact and terminal events	Reference-state error; rule assertions; controlled replay
Persistent state	Identity and consequences after occlusion, revisit, or delayed events	Revisit/mutation probes; save–restore; synchronized views

Table 7: Complementary tests of world-model fidelity. State-based checks require an independent specification or reference; visible agreement must also be checked when state drives a learned renderer.

Model	GC	IF	IE	OE	Overall	Validity (%)
Genie 3	2.74	2.40	1.51	1.81	2.12	87.1
HappyOyster	2.54	2.15	1.47	1.54	1.92	79.6
LingBot-World2	2.04	2.13	1.95	1.16	1.82	78.8
LingBot-World	2.11	2.23	1.33	1.43	1.78	72.7
HY-World2	2.14	2.06	1.13	1.09	1.61	51.7
SANA-WM	1.72	1.89	1.13	1.16	1.48	80.4

Table 8: PlayWorld evaluation of nine interactive world models ( Ding et al., 2026 ) . Rubric scores range from 1 to 5; trajectory validity is a percentage. Higher values are better in every column.

Model (effort, where specified)	Harness	Pass@1 (%) $\uparrow$
Claude Fable 5 (xhigh)	Claude Code	$67.3\pm 5.0$
GPT-5.6 Sol (xhigh)	Codex	$63.7\pm 5.2$
GPT-5.6 Sol (high)	Codex	$63.1\pm 5.2$
Muse Spark 1.2 (high)	Muse Code	$61.0\pm 5.2$
GPT-5.6 Sol (medium)	Codex	$58.6\pm 5.3$
Kimi K3	Kimi Code	$58.0\pm 5.3$

Table 9: GameDevBench results for 17 model–harness configurations evaluated on 333 Godot tasks ( Chi et al., 2026 ; GameDevBench Team, 2026 ) . Values are Pass@1 estimates with 95% confidence intervals.

Model (effort)	Harness	Core mechanics	Content depth	Functional visuals	Art and presentation	Overall
Claude Opus 5 (xhigh)	Claude Code	76.75	63.31	68.04	70.11	68.44
Claude Fable 5 (high)	Claude Code	76.52	58.61	66.93	67.68	65.72
GPT-5.6 Sol (high)	Codex	74.50	56.10	64.80	57.00	60.50
Kimi K3	Kimi Code	73.03	53.38	56.90	53.59	56.96
Seele02-pro	SeeleAgent	68.42	48.76	52.74	44.17	50.70
Claude Opus 4.7 (high)	Claude Code	55.34	39.48	42.78	36.86	41.46

Table 10: GameCraft-Bench results for 14 model–harness configurations across 140 tasks in 15 game families ( Luo et al., 2026 ; GameCraft-Bench Team, 2026 ) . Scores are normalized to a 0–100 scale, not project pass rates; higher values are better.

Held out	What changes	Claim supported by testing	Example / test
Map / level	Layout, placement, visual setting	Robustness within specified rules	Procgen: unseen levels [ 66 ]
Mode	Objectives, opponents, difficulty, or rules	Adaptation to named rule/task variants	Atari mode transfer [ 97 ]
Game	Title; potentially rules, assets, controls	Transfer of the evaluated behavior or model function	SIMA: task success; SCOPE: action response [ 291 , 328 ]

Table 11: Map, mode, and game changes test different generalization settings. The measured capability and permitted target adaptation must be specified in each case.

Benchmark / setting	Task and measure	Evidence	Comparison limits
Play and Act
TextWorld / Jericho / TextArena [ 74 , 126 , 115 ]	Generated quests, authored adventures, and single-/multi-player text games; task-specific outcomes	Play; executable checks	Command vocabulary, rule access, partners, and scoring conventions differ
NLE / Crafter [ 164 , 121 ]	Procedural exploration and survival; game scores and achievement coverage	Play; executable checks	Progress within a game, not arbitrary-rule transfer
BALROG [ 239 ]	Six game environments; text/visual inputs; normalized progress	Play; executable checks	Abstracted actions; native motor timing excluded
VideoGameBench [ 394 ]	Full: 10 real-time test games; Lite: three development and three test games, with the game paused during model inference; checkpoint progress	Gameplay traces	Different game sets; progress is not completion
GameWorld [ 238 ]	170 tasks / 34 browser games; state-verifiable success and partial progress	Interaction; state checks	Native vs. semantic actions; game paused during inference

Table 12: Benchmarks and evaluation settings across the six roles: measured targets, principal evidence sources, and limits of interpretation.

Flow	Artifact	Example system	Outcome to validate
Play $\rightarrow$ Model	Action-linked traces	GameNGen [ 333 ]	Prediction fidelity; interactive consistency
Model $\rightarrow$ Play	Imagined trajectories	Dreamer 4 [ 125 ]	Reference-game policy performance
Design $\rightarrow$ Build	Editable plan	DreamGarden [ 90 ]	Executable output satisfying requirements
Build $\rightarrow$ Model	Executable state; rendering controls	Programmable World Model [ 143 ]	Generated views agreeing with engine events
Test $\rightarrow$ Design	Compiler/search feedback	ScriptDoctor [ 89 ]	Compilation; solvability within search budget
Design $\rightarrow$ Runtime	Narrative structure	NarrativeGenie [ 163 ]	Narrative continuity; intended player experience

Table 13: Representative implemented cross-role exchanges and downstream outcomes to validate.

Role	Reported evidence	Evaluation conditions	Research gap
Play	Cross-game skills [ 292 , 206 ] ; rule inference [ 295 ]	Selected games; specified adaptation budgets	Joint transfer across rules, controls, and timing
Model	State prediction [ 193 ] ; NPC behavior [ 79 ] ; player prediction [ 319 ]	Domain-specific state and behavior data	Persistent world state; reliable player-model updates and transfer
Design	Solvable levels [ 302 ] ; rule search [ 326 ] ; designer steering [ 90 ]	Fixed representations; solver/creator feedback	Intended player experience; sustained creative control
Build	Complete projects [ 202 ] ; scoped edits [ 165 ] ; short revision chains [ 56 ]	Specified engines, requests, tests	Evolving requirements; developer handoff
Runtime	Constrained rule execution [ 134 ] ; player-outcome studies [ 133 , 120 ]	Bounded encounters or field deployments	Cross-session consistency, experience, and shared-state fairness
Test	Mechanic verification [ 149 ] ; human-reference defect testing [ 19 ] ; fairness diagnostics [ 285 ]	Specified access, oracles, sampled behaviors	Independent verification; broader behavioral coverage

Table 14: Representative evidence and research gaps across the six roles. Findings are setting-dependent, not a ranking of role maturity.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

System / source	Role	Environment and mechanism	Data and evaluation
Shared policies and learned controllers
VPT [ 24 ]	P	Env: Minecraft; pixels; keyboard/mouse Mech: Inverse-dynamics labels; BC, then RL	Data: About 70K h of video; smaller labeled set Eval: Native control; Minecraft-specific pretraining
Gato [ 263 ]	P	Env: Atari plus language/robotics tasks Mech: One autoregressive, multimodal policy	Data: Mixed offline demonstrations Eval: Shared weights across tasks; adaptation evaluated separately
Multi-Game DT [ 171 ]	P	Env: Atari; pixels and discrete actions Mech: Offline return-conditioned Transformer	Data: 41 training / 5 held-out games Eval: Shared training-game policy; new titles require fine-tuning
DreamerV3 [ 124 ]	P ( M )	Env: Visual or vector observations Mech: Latent world model; imagination RL	Data: 150+ tasks; one configuration Eval: Algorithm reuse with per-task learning
STEVE-1 [ 189 ]	P	Env: Minecraft; language goals; native actions Mech: VPT policy; latent goals; hindsight	Data: Relabeled Minecraft behavior Eval: Text-to-behavior transfer within one game

Appendix

Table 15: Game-playing systems and their transfer conditions. Grouped entries share a technical purpose, not a common score scale.

System / source	Role	Representation and control	Data and evidence
Planning and policy-learning models
World Models [ 119 ]	M ( P )	Repr: VAE + recurrent latent dynamics Control: Continuous control; visual observations	Data: CarRacing and VizDoom Evidence: Controller return; transfer from imagined to real rollouts
MuZero [ 275 ]	P ( M )	Repr: Value-, reward-, and policy-trained latent state Control: Discrete actions; tree search	Data: Atari, Go, chess, shogi Evidence: Decision quality; no image reconstruction requirement
Dreamer / DreamerV3 [ 123 , 124 ]	P ( M )	Repr: Recurrent latent dynamics; imagined trajectories Control: Actions and rewards	Data: Visual and vector control tasks Evidence: Policy return in reference environments
I2A [ 260 ]	P ( M )	Repr: Encoded rollouts from an imperfect model Control: Discrete actions	Data: Sokoban; MiniPacman Evidence: Policy use of prediction errors
Action-conditional prediction [ 235 ]	M	Repr: Feedforward / recurrent image prediction Control: Recorded Atari actions	Data: Five Atari games Evidence: Prediction error; selected control tests

Appendix

Table 16: Learned game models: what they represent, how they are controlled, and what is evaluated. Reported speed and horizon use different hardware and protocols.

System / source	Role	Input and construct	Validation and boundary
Behavior, skill, and individual action prediction
player2vec [ 343 ]	M	Input: Mobile-game event sequences Model: Masked-token Longformer	Data: Masked-event recovery; embedding structure Boundary: No independent preference labels
Behavior Structformer [ 296 ]	M	Input: Structured tracking events Model: Transformer over typed event fields	Data: Constructed session-count targets Boundary: Activity proxy rather than experience
TrueSkill [ 129 ]	M	Input: Match outcomes Model: Bayesian skill and uncertainty	Data: Predictive performance of ratings Boundary: Historical skill model; not preference inference
Maia / individual Maia [ 214 , 215 ]	M	Input: Chess positions and player histories Model: Skill-group / individual behavior models	Data: Held-out human moves Boundary: Behavior imitation, not strongest play
Maia-2 [ 318 ]	M	Input: Position and skill level Model: Shared model; skill-aware attention	Data: Human moves across rating groups Boundary: Population conditioning

Appendix

Table 17: Player modeling by prediction target. Input, inferred construct, and intervention outcome occupy different columns; datasets and frameworks are marked with a dash.

System / source	Role	Artifact and mechanism	Feedback and evaluation
Content, rules, and mixed-initiative design
Tanagra / Sentient Sketchbook [ 299 , 187 ]	D	Artifact: Platformer levels / map sketches Mech: Constraint solving / novelty search	Feedback: Playability checks; designer edits Eval: Historical mixed-initiative authoring
ANGELINA / Mechanic Miner [ 73 , 72 ]	D	Artifact: Games and mechanics Mech: Automated design; reflective mechanism search	Feedback: Simulated play and design criteria Eval: Historical automated-game-design systems
CubePart [ 416 ]	D	Artifact: Part-controllable 3D assets Mech: Semantic schemas; separate meshes	Feedback: Part identity and geometry Eval: Part control; external behavior scripts
WorldSculpt [ 231 ]	D	Artifact: Compositional mesh scenes Mech: Posed views; object mesh prior	Feedback: Mesh/scene fidelity Eval: Unreal reconstruction tests
AutoBG [ 186 ]	D ( T )	Artifact: Board-game ideas and rulebooks Mech: Rule iteration; persona feedback	Feedback: Designer revision; model critiques Eval: 207 held-out games; 30-creator study

Appendix

Table 18: Design and development: proposed artifacts, operating mechanisms, feedback, and evaluation settings. Benchmark rows have no AI-role assignment.

System / source	Role	Generated element and execution	Human evidence and finding
Narrative and content systems
CALYPSO [ 411 ]	R ( D )	Element: Encounter prose; summaries; ideas Execution: Human-mediated D&D; current scenario	Exposure: 71 players and DMs; field deployment Finding: 51 chat threads; direct use plus DM revision
Concordia [ 335 ]	R ( P )	Element: Character actions and narrated consequences Execution: Actor / game-master components; memory	Exposure: Framework demonstrations Finding: Configurable social simulation; no common player-outcome trial
KG-conditioned quests [ 21 ]	R ( D )	Element: Personalized quests and dialogue Execution: Player input; graph paths; grammar; GPT-2	Exposure: Text-rating study; generated-quest analysis Finding: Intent grounding; unsupported entities remain possible
NarrativeGenie [ 163 ]	D ( R )	Element: Story graphs; hints; summaries Execution: Graph/state dialogue manager	Exposure: 19 players; one mystery episode Finding: Ratings plus 30 simulated-history checks
Drama Llama [ 307 ]	R ( D )	Element: Responsive text narrative Execution: Storylet triggers checked by an LLM	Exposure: Six authors; authoring study Finding: Authorable responses; not a player-outcome trial

Appendix

Table 19: Runtime systems: generated elements, execution context, and human evidence. Raters, authors, and interactive players are distinguished; n.s. means no statistically significant difference was reported.

System / source	Role	Target and tester	Oracle and evidence
Exploration, coverage, and regression
Procedural Personas [ 130 ]	T	Target: Executable game states Tester: MCTS; evolved goal heuristics	Oracle: Game-specific outcomes Evidence: Deliberate style diversity; not population sampling
Wuji [ 407 ]	T	Target: Instrumented combat games Tester: Multiobjective evolutionary RL	Oracle: Bug predicates; developer confirmation Evidence: Two commercial games; three newly confirmed bugs
Synthetic / humanlike agents [ 19 ]	T	Target: Three GVG-AI games Tester: Agents with different behavior objectives	Oracle: Separate execution oracle; human trajectories Evidence: 45 seeded bugs; 427 human traces
EA navigation AutoPlayers [ 108 ]	T	Target: Battlefield 2042 / Dead Space; structured state Tester: RL navigation within scripted test bots	Oracle: Existing production QA infrastructure Evidence: Navigation deployment, not end-to-end learned QA
CA2 [ 3 ]	T	Target: State/images and call stacks Tester: RL guided toward code targets	Oracle: Function-entry instrumentation Evidence: Improved target reachability; diagnosis is separate

Appendix

Table 20: AI testing systems: exploration, oracles, and evaluation. Named testing components are indexed separately.

Game / system	Family	Role	AI contribution	Resource / status
Players, opponents, and teammates
Minecraft / Voyager [ 341 ]	Sandbox	P	Code-skill acquisition	Research: code and videos
GT7 / Sophy † [ 356 , 257 ]	Racing	P	Learned rival driving	PS5 custom-race feature
PUBG / Ally [ 161 , 294 ]	Battle royale	P , R	Voice-guided cooperation	Limited beta: June 2026
StarCraft II / AlphaStar † [ 337 ]	RT strategy	P	Strategy and unit control	Research match recordings
Dota 2 / OpenAI Five † [ 33 ]	Team MOBA	P	Team coordination	Historical match videos

Appendix

Table 21: Game examples and public materials. P: Play; M: Model; D: Design; B: Build; R: Runtime; T: Test.

\dagger

: specialist or pre-foundation-model lineage.

Automated Game Testing

See all 10 papers

May 11, 2026

Towards Generalist Game Agents: A Survey of Foundation Models in Virtual Worlds

The real world unfolds along a single set of physics laws, yet human intelligence demonstrates a remarkable capacity to generalize experiences from this singular physical existence into a multiverse of games, each governed by entirely different rules, aesthetics, physics, and objectives. This omni-reality adaptability is a hallmark of general intelligence. As Artificial Intelligence progresses towards Artificial General Intelligence, the multiverse of games has evolved from mere entertainment into the ultimate ground for training and evaluating AGI. The pursuit of this generality has unfolded across four eras: from environment-specific symbolic and reinforcement learning agents, to current large foundation models acting as generalist players, and toward a future creator stage where agent both creates new game worlds and continually evolves within them. We trace the full lifecycle of a generalist game player along four interdependent pillars: Dataset, Model, Harness, and Benchmark. Every advance across these pillars can be read as an attempt to break one of five fundamental trade-offs that currently bound the whole system. Building on this end-to-end view, we chart a five-step roadmap, progressing from single-game mastery to the ultimate creator stage in which the agent simultaneously creates and evolves within theoretical game multiverse. Taken together, our work offers a unified lens onto a rapidly shifting field, and a principled path toward the omnipotent generalist agent capable of seamlessly mastering any challenge within the multiverse of games, thereby paving the way for AGI.

Kuan Zhang, Dongchen Liu, Qiyue Zhao +12

College of AI, Tsinghua University · MMLab, The University of Hong Kong · University of Chinese Academy of Sciences

Game-Playing Agents AI Agent Benchmarks

Ask PDF

Sep 21, 2026

GameHorizon Suite: Multi-Horizon Data and Evaluation in Gameplay

Modern video games provide a measurable testbed for AI models, combining abilities of visual understanding, instruction decomposition, goal planning, and precise action control over multiple temporal horizons. Existing datasets and benchmarks, however, either cover a narrow range of games, lack language instructions, or rely on high-variance online rollouts. To address these challenges, we introduce GameHorizon, a unified data and evaluation suite that measures gameplay capabilities at different horizons for diverse model families. GameHorizon Suite consists of three components. First, GameHorizon-Annotator is a scalable and automated annotation pipeline for multi-horizon instructions. Second, utilizing the pipeline, we construct GameHorizon-Data, the first large-scale AAA gameplay dataset with temporally aligned videos, player actions, and multi-horizon instructions. It comprises 5,000 hours of recordings from 21 games, collected by 100 human expert players. Third, we build GameHorizon-Bench with reproducible offline and stepwise online testing. The offline track enables reproducible evaluation using thousands of standardized questions organized into three primary tasks and a series of diagnostic variants, while the online track tests whether offline scores reflect actual gameplay capabilities and localizes failures to specific steps within long-horizon gameplay. Based on our GameHorizon Suite, we evaluate 47 models through more than one million model invocations, revealing a meaningful hierarchy of task difficulty and pronounced differences in model capabilities. Our work can provide a standardized yardstick for evaluating gameplay capabilities across horizons and model families. We will release our dataset, annotator, and benchmark to facilitate future research.

Yiran Wang, Xingyilang Yin, Junfu Pu +12

ARC Lab, Tencent · GVC Lab, Great Bay University · University of Macau +3

Game-Playing Agents Long-Horizon Agent Evaluation Agent Benchmarks

Ask PDF

Aug 14, 2026

Twin: Playing an Unknown Game with a Test-Time Digital Twin

We present a Test-time World-model Inference (Twin) system, in which a frontier coding agent writes an executable world model for completing continual learning tasks, such as ARC-AGI-3 games. Traditional approaches hand-engineer such models, one custom design per task. Each game hides its rules and goal, and our system constructs them from simulation and interaction alone. Its inductive prior over grid games is strong enough to recover the true transitions of the game and the goal on nearly all levels. Replay validation happens in a twin world model. The harness enforces that an action is not made until the program reproduces every previous observed game transition. Each mismatch between a world model prediction and the actual action result becomes a counterexample that is used to repair the world model. Twin clears 179 out of 183 levels (97.8%), and does so more efficiently than humans in 158 out of 179 levels (88.3%). The system infers the goal before any reward on 156 of the levels it clears (87.2%), and in the remaining levels automatically discovers the goal by search. The benchmark scores completion and action efficiency, between 0 and 100, against humans playing each game for the first time. Played directly, the base model scores only 7.8%; an off-the-shelf harness increases it to 61.1%, whereas our twin world model increases the same base model to 93.3%, clearing 23 out of 25 games. Building a usable world model is simpler than anticipated, whereas the harder problem is inferring the right goal.

Alexy Skoutnev, Kirill Acharya, Gaston Longhitano +3

Yeshiva University · Stanford University · University of Southern California +1

Game-Playing Agents Code World Models World Model Learning ARC-AGI Digital Twins

Ask PDF

Abstract

Automated Game Testing

Explore similar work

Towards Generalist Game Agents: A Survey of Foundation Models in Virtual Worlds

GameHorizon Suite: Multi-Horizon Data and Evaluation in Gameplay

Twin: Playing an Unknown Game with a Test-Time Digital Twin