AI for Games in the Foundation Model Era - Comprehensive Summary

Summary (Overview)

• Six-Role Taxonomy: The paper organizes AI-for-games research into six roles based on the immediate use of AI output: (1) Play and Act, (2) Model Players and Games, (3) Design, (4) Build and Maintain, (5) Generate and Adapt at Runtime, and (6) Test and Evaluate.
• Cross-Role Connections: The survey identifies concrete exchanges between roles—trajectories train world models, learned environments provide experience for agents, design specifications drive implementations, and play/test feedback guides revision—while emphasizing that artifact reuse does not automatically establish capability transfer.
• Three Recurring Findings: Broad pretraining expands available interfaces without removing game-specific structure; gameplay increasingly supplies data beyond an agent's final score; progress remains task-dependent with stronger benchmarks for bounded play than for sustained creation, adaptation, and human experience.
• Evidence Assessment: 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.
• Central Challenge: Enabling outputs and capabilities to be reused across roles while re-establishing evidence for their effectiveness in game-specific structures, interfaces, and player contexts where they are ultimately used.

Introduction and Theoretical Foundation

Background and Motivation

AI for games has historically extended beyond playing. The paper cites recent applications of GPT-6 Astra in both unfamiliar interactive environments (ARC-AGI-3) and game development tools (Playco's Playbot) to illustrate how the same pretrained model can serve different roles across the game lifecycle. The survey connects several established research traditions:

• Game-playing research: Methods for selecting actions under game rules (Shannon, 1950; Mnih et al., 2015; Silver et al., 2016)
• Procedural generation: Producing game content (Togelius et al., 2011) and automated design (Browne and Maire, 2010)
• Interactive narrative: Shaping stories in response to player actions (Mateas and Stern, 2005)
• Player modeling: Predicting preferences to guide content adaptation (Yannakakis and Togelius, 2011)
• Automated playtesting: Simulated players examining game aspects (Holmgård et al., 2019)

Organizing Framework

The survey organizes AI for games by the immediate use of the system's output:

Three Recurring Questions

1. Boundary: What is supplied by the game or workflow, and what is assigned to AI?
2. Transfer and Reuse: Which capabilities transfer, which artifacts can be reused, and what remains setting-specific?
3. Evidence: What claims are supported by evaluation at the point of use?

Methodology

Taxonomy Assignment

Systems are classified by primary role based on the immediate use of their output, with cross-references recording secondary roles. For example:

• An NPC policy belongs to Play and Act when the claim concerns action quality
• Session-specific dialogue evaluated for player experience belongs to the runtime role
• A playtester belongs to Test and Evaluate when the claim concerns coverage or defects

Analytical Approach

For each role, the survey examines:

1. What structure is supplied by the game or workflow
2. What AI learns, generates, predicts, or revises
3. Which capabilities transfer and which artifacts can be reused
4. What claims are supported by available evidence

Cross-Role Framework

The paper proposes five dimensions for comparing evidence across roles:

• Standardization: Are tasks, protocols, and measures shared across studies?
• Execution grounding: Is the claim checked against executable behavior?
• Scope and transfer: Does evidence span games, genres, engines, or player populations?
• Horizon and revision: Does evaluation cover sustained interaction or repeated edits?
• External validation: Is the reference independent and representative?

Empirical Validation / Results

Play and Act

Specialist agents (AlphaStar, OpenAI Five, Gran Turismo Sophy) establish expertise within single games but do not demonstrate transfer. Algorithm reuse (General Game Playing, DreamerV3, Procgen) tests whether learning procedures transfer while policies are retrained. Parameter sharing (Gato, SIMA 2, NitroGen) retains trained weights across games.

Key benchmark results:

Table 5: GameWorld Results (Ouyang et al., 2026) - 170 tasks in 34 games

Table 6: ARC-AGI-3 Results - GPT-6 Astra scores 54.82 (Standard harness) vs. 99.95 (Provider Adapter) at high reasoning effort, demonstrating significant harness sensitivity.

Model Players and Games

World-model fidelity has three distinct targets:

• Perceptual quality: FVD, perceptual scores, human ratings
• Mechanics correctness: Reference-state error, rule assertions, controlled replay
• Persistent state: Revisit/mutation probes, save-restore, synchronized views

PlayWorld evaluation of nine interactive world models (Ding et al., 2026):

StatePlay (Street Fighter III): normalized state error below 0.06, +18.6 percentage points over stateless baseline on mechanics-fidelity judgments.

Maia4All: Held-out move accuracy rises from ~51.4% (Maia-2) to 53.2% with as few as 800 positions (~20 games) of individual player data.

Design

MarioGPT: 88.4% of 250 generated levels solved by A* agent; separate 1,000-sample experiment measures prompt adherence (pipes, enemies, blocks, elevation).

GAVEL: Fine-tuned code model mutates Ludii descriptions within quality-diversity search; compilation and simulated play screen proposals.

DreamCraft: Minecraft block-aligned 3D generation with quantized neural radiance fields, improving in-game text alignment over post-hoc conversion.

Sokoban experiments (Todd et al.): With abundant Boxoban data, pretraining brings little benefit; controlling empty-space proportion is easier than controlling solution length.

Build and Maintain

Table 9: GameDevBench Results - 333 Godot tasks, Pass@1 estimates:

Table 10: GameCraft-Bench Results - 140 tasks, 15 game families:

GameEngineBench: Best configuration solves 55.5% of 110 tasks; 31 tasks defeat every tested configuration. Many failures compile successfully but violate engine lifecycle, replication, or cross-system behavior.

JAMER: Runtime-pass rates drop from 80.4% (small projects) to 5.7% (large projects) for function-body reconstruction.

Generate and Adapt at Runtime

IF:CARGO (24 participants, 8 levels): Players used execution feedback to revise rules; periodic commands, multi-robot coordination, and priority rules increased diagnostic demands.

GenFlora (72 participants, 2×2 within-subject): Both dynamic item functionality and responsive NPC dialogue improve presence, autonomy, and enjoyment independently.

Roso et al. (34 participants): Sustained token delay has more pronounced negative effects than initial wait in RPG conversations.

Hafnar & Demšar (102 players, 928 level starts): 55% completion under LLM condition vs. 35% with random parameter selection, but first-level completion also differed before personalization history was available.

Test and Evaluate

GameGen-Verifier: 92.2% specification-label agreement with expert annotations (Acc@5), vs. 58.8% for coverage-enforced baseline, up to 16.6× lower wall-clock time.

VideoGameQA-Bench: Best visual-regression accuracy is 45.2%, despite stronger glitch-recognition results.

TITAN: 95% completion on 20 tasks across two commercial MMORPGs, four previously unknown bugs, adoption in eight QA pipelines.

PlaytestArena: 84.2% raw criterion agreement with human judgments on 32 games (Cohen's κ = 0.64).

Theoretical and Practical Implications

Cross-Role Connections

The paper identifies concrete implemented exchanges:

Key Insights

1. Broader interfaces do not remove game-specific structure: Language, multimodal, and code interfaces let AI participate in more workflow stages, but the connection to executable behavior differs across systems. What the model can express and what the game can execute remain distinct.

2. Artifact reuse ≠ capability transfer: A shared backbone or successful component communication does not demonstrate transfer. Transfer claims require specifying what changes and testing whether relevant competence survives.

3. Feedback loops create blind spots: When a judge guides repeated revision, its blind spots can be reinforced rather than revealed. Independent execution checks and withheld defects matter beyond ordinary judge agreement.

4. Diversity is not representativeness: Procedural personas broaden test behavior, but human-trajectory comparisons are needed when the claim is representativeness rather than diversity alone.

Emerging Direction: Executable Grayboxes and Generative Rendering

The paper identifies a concrete emerging connection separating game construction, state execution, and visual synthesis:

• A code agent translates intent into an executable graybox (geometry, controls, collision, objectives, tests)
• A state-conditioned video model synthesizes appearance from engine buffers
• Input returns to the engine, so generated pixels do not silently redefine collision or score

This full production-engine combination remains a research direction; the lightweight-engine demonstration does not establish reliable UE/Unity project generation workflows.

Conclusion

Main Takeaways

1. Foundation models expand what AI can interpret, generate, predict, and revise across the game lifecycle, but dependence on explicit game structure (controls, rules, state representations, engine interfaces, human decisions) persists.

2. The most concrete cross-role connections arise through reuse of interaction traces, learned environments, specifications and rules, execution/test evidence, and player models—but artifact reuse does not by itself establish capability transfer.

3. Evidence is strongest and most standardized for bounded game playing and selected learned environments; persistent world state, evolving software projects, repeated revision and handoff, sustained runtime adaptation, and repeated player encounters remain less established.

Future Directions

1. Play and Act: Joint transfer across rules, controls, and timing; executable rule acquisition connected to real-time control; factorial tests separating map/mode/game changes.

2. Model Players and Games: Persistent world state with selectively explicit, updatable state; reliable player-model updates distinguishing stable tendencies from session-specific state.

3. Design: Separating hard constraints from revisable preferences; preserving diverse playable alternatives; extending control to language- and code-based generators.

4. Build and Maintain: Version-aware project context retaining accepted requirements and reproducible tests; multi-version benchmarks; developer handoff evaluation.

5. Generate and Adapt at Runtime: Cross-session consistency with conservative fallbacks; isolating adaptation contributions from generation quality; repeated-session trials measuring continuity and agency.

6. Test and Evaluate: Combining complementary exploration policies with independent oracles; withheld defect families; human relevance sampling across novice/expert trajectories and fairness cases.

Central Challenge

"The central challenge, therefore, is not simply to move outputs across the game lifecycle, but to preserve the information and constraints that make those outputs useful while re-establishing evidence in the settings where they are ultimately applied. Progress in AI for games will depend on extending reuse and transfer without losing developer control, game consistency, or relevance to the players these systems are intended to serve."