Large language model agents have enabled software development through natural-language interaction, but extending this paradigm to end-to-end game development remains challenging. Building a playable game requires coordinating code, scenes, assets, animations, physics, and configurations, while allowing developers to validate their integrated behavior through interactive testing. We introduce VibeGame, a framework for creating complete 2D game projects from natural-language prompts and optional reference images. Its AI-native game engine represents project state as schema-validated structured text and provides frame-synchronous runtime control, enabling agents to inspect game states, execute actions, and verify behavior at their own pace. Built upon this interface, we design an Adversarial Agent Team of eight specialized agents to separate implementation from review, so that problems are found and fixed through multi-round loops. For each conducted and accepted project, we further introduce a training-free self-evolution mechanism to distill it into reusable skeletons, modules, and collaboration contracts for future development. We evaluate prompt-assets-to-game (PA2G), where agents are given an asset library, and prompt-to-game (P2G), where agents generate their own art. On a GameCraft-Bench subset, VibeGame in P2G consistently outperforms a range of agent harnesses in either P2G or PA2G with the same underlying model, including GPT-6.1-Sol.
Figures & tables
Figure 2: Games created by VibeGame from prompts across diverse 2D game genres.
Figure 3: System overview of VibeGame. (a) The Adversarial Agent Team turns a user request into an accepted game through intention alignment, parallel development, and adversarial correction. (b) The AI-native game engine exposes a structured project representation and frame-synchronous runtime control for implementation and play-testing. (c) Training-free self-evolution extracts skeletons, modules, and collaboration contracts from accepted projects and returns them to future development as project priors.
Harness
Model
Input
M
D
V
A
Overall
OpenGame [ 9 ]
GPT-5.6-Luna
P2G
36.78
25.57
23.30
44.57
33.58
Qwen Code [ 28 ]
GPT-5.6-Luna
PA2G
46.38
32.18
27.88
37.15
35.45
Kimi Code [ 29 ]
GPT-5.6-Luna
P2G
45.26
32.13
31.00
38.34
36.14
Kimi Code
GPT-5.6-Luna
PA2G
47.54
31.86
29.67
38.46
36.22
Codex [ 7 ]
GPT-5.6-Luna
PA2G
50.72
33.32
28.33
36.87
36.47
Codex
GPT-5.6-Luna
P2G
45.22
34.47
31.54
37.65
36.77
Table 1: Results on the GameCraft-Bench subset. DSH denotes DeepSeek Harness. M, D, V, and A denote Core Mechanics, Content Depth, Functional Visuals, and Art & Presentation, respectively; Overall is the weighted score defined in Appendix 6 . Bold numbers in green cells and underlined numbers in amber cells denote the best and second-best scores per column within each model, respectively.
Figure 4: Results by game type and pairwise win rates. (a) Gain of VibeGame over the strongest baseline of each model, Codex (P2G) for GPT-6.1-Sol and Qwen Code (P2G) for GPT-5.6-Luna, on each game type. (b) Win rate of VibeGame against each GPT-6.1-Sol baseline in the user study (top) and under the GPT-6.1-Sol judge (bottom); each bar stacks the win rates in the four rubric categories (M, D, V, A), and the number at its end is the overall win rate.
Figure 5: Qualitative comparison between VibeGame, Codex (PA2G), and Codex (P2G), all using GPT-6.1-Sol. Left: Sports Skateboard Park. Right: Sports Boxing Gym. Red boxes mark the defects.
Variant
M
D
V
A
Overall
VibeGame
60.73
48.58
43.10
59.26
53.45
w/o Engine
52.16
32.95
32.98
53.22
43.12
w/o AAT
57.79
41.94
34.43
58.09
48.82
w/ Experience
68.76
47.47
40.42
66.39
56.37
Table 2: Ablation on the ablation set (GPT-5.6-Luna). Columns and marks as in Table 1 .
Harness
Model
Input
M
D
V
A
Overall
OpenGame
GPT-5.6-Luna
P2G
46.04
33.90
33.17
47.22
40.29
Codex
GPT-5.6-Luna
P2G
58.51
45.23
48.69
47.24
48.46
Kimi Code
GPT-5.6-Luna
P2G
61.04
43.42
52.14
47.67
48.89
Qwen Code
GPT-5.6-Luna
PA2G
66.67
46.47
51.61
49.93
51.51
Kimi Code
GPT-5.6-Luna
PA2G
65.79
44.68
50.50
52.73
51.57
Qwen Code
GPT-5.6-Luna
P2G
64.56
47.71
53.30
50.17
51.97
Table 3: Results on the GameCraft-Bench subset scored by the GPT-5.5 judge (mean of its passes). Rows, columns, and marks as in Table 1 .
Figure 6: Judge error case 1: Kimi Code + GPT-5.6-Luna (P2G) on Racing Drift Circuit. The judged track-select frame: its twelve thumbnails reuse four layouts, one per biome (e.g., the three Canyon tracks share one layout).
Figure 7: Judge error case 2: OpenGame + GPT-5.6-Luna on Cardgame Autobattler. The board has a front and a back row (BACK 2–4) and no bench; the board heading and row labels are covered by cards.
Figure 8: Judge error case 3: Kimi Code + GPT-5.6-Luna (P2G) on Idle Factory Planet. Three consecutive judged frames, cropped to one production chain and shown at the source recording’s 1280 × 720 resolution (the judges received them downscaled to 854 × 480): conveyors connect the four machines and the resource particles move along them.
Game type
Tasks
Card Game
Autobattler † , Poker Roguelike †
Horror
Lighthouse † , Tape Archive
Idle
Dungeon Guild † , Factory Planet
Open World
Beast Tamer, Sky Islands
Platformer
Shape Shift, Thunder Valkyrie
Puzzle
Gravity Shift † , Pipe Crisis †
Table 4: The 30 benchmark tasks, two per game type. † marks the 10 tasks of the ablation set.
Figure 9: Game creation demonstrations. Top: Pixel Sekiro, a pixel-art boss fight created from a text prompt. Bottom: Hollow Knight boss slice, a side-scrolling boss fight created from a gameplay prompt and a concept image.
Figure 10: Game edit demonstrations. Top: intellectual property (IP) transfer, replacing the visual theme while preserving gameplay logic. Middle: genre transfer, converting a turn-based card game into a real-time boss fight while preserving its visual identity. Bottom: rule transfer, swapping the player and boss roles while preserving their abilities and animations.
Figure 11: A three-fruit combo sliced mid-air with split halves and juice effects.
Figure 12: Two fighters trade blows with an impact spark on the neon stage.
Figure 13: The player fires a spread of projectiles at an enemy inside a sealed combat room.
Figure 14: The protagonist floats above a balloon, with the goal flag on its platform.
Figure 15: A selected warrior unit and its stats beside a city on the hex map.
Figure 16: The round-table layout with AI player portraits and live speech text during the sheriff election.
Figure 17: Tanjiro faces Upper Rank Three under the moon on the Mugen Train night field.
Figure 18: Auditor case, before (top) and after (bottom). After the repair, each band is placed and sized by its tapping window on one time axis. Placeholder art; the bottom image shows the physics debug overlay.
Figure 19: Player case, before (top) and after (bottom). After the repair, every entry shows its full name, type, and status.
Figure 20: Reviewer case, before (top) and after (bottom). After the repair, the companion stands beside the keeper.
Harness
M
D
V
A
Overall
Codex
80.21
58.42
57.51
66.52
64.44
w/ Engine
82.41
62.37
58.94
68.75
67.13
Table 5: Codex with and without the AI-native engine of VibeGame (GPT-6.1-Sol, PA2G, all 30 tasks). Columns as in Table 1 .
Recent advances in large language models have made automatic game generation increasingly feasible, yet reliably improving generated games beyond a playable version remains challenging. Naive iterative refinement can easily overfit a small set of test cases, producing fragile games with unresolved bugs, missing behaviors, and poor generalization to broader player interactions. We introduce RSIGame, an autonomous agentic game development framework with recursive self-improvement. RSIGame organizes development into complementary local and global loops. Concretely, a local explore-diagnose-improve loop broadly explores the executable game, diagnoses and prioritizes discovered issues, and performs evidence-grounded revision, where an evolving checklist continually accumulates new testing and improvement guidance. A global loop tracks overall quality, preserves the best checkpoint, and detects saturation or regression over long-horizon development. Beyond test-time improvement, RSIGame further internalizes successful development experience into the generator through training. Across 140 GameCraft-Bench tasks, two game engines, and five generators, RSIGame consistently improves game quality under matched development budgets. Notably, experience internalization enables Qwen3.8-27B to reach 61.38 on Godot and 58.53 on Phaser, exceeding GPT-5.5 one-shot scores while reducing Qwen's generation tokens by 11 times.
Wenyi Wu, Minghao Fu, Jieyu You +10
University of California San Diego · ByteDance Inc. · Carnegie Mellon University
Recent advances in Large Language Models (LLMs) have demonstrated remarkable capabilities in web front-end execution, with browser-based game generation emerging as a particularly prominent frontier. While previous efforts frequently rely on complex multi-turn workflows or focus on static game evaluation benchmarks, this work targets direct end-to-end real-world game synthesis driven by coding agents. However, generating complex games directly from sparse user queries often forces coding agents to make underspecified assumptions, yielding incomplete mechanics, disconnected gameplay flows, and limited visual aesthetics. To resolve this issue, this paper presents GameGo, a scalable framework that systematically transforms brief game seeds into comprehensive Product Requirements Documents grounded in industry game-development practices. To retain core gameplay constraints without restricting design exploration, GameGo uses task-specific dynamic compression to maximize information density while preserving instruction following. Based on this pipeline, GameGoData is constructed with 55,060 development trajectories across 2D, 2.5D, and 3D games, alongside GameGoBench, a benchmark comprising 124 diverse game queries. Training GameGoCoder on GameGoData yields a model that outperforms matched baselines and is comparable to frontier models across gamedev benchmarks. All code, datasets, and models will be made publicly available.
Haoyue Yang, Jingyao Li, Zhengfan Wu +3
Institute of Automation, Chinese Academy of Sciences · Baidu Inc., Beijing, China
Game generation is an emerging application of coding agents, requiring models to transform natural-language specifications into playable interactive systems. Unlike traditional coding tasks, game generation takes place within a game engine, where scripts, scenes, assets, rendering, and runtime interactions must jointly produce coherent gameplay. We formalize end-to-end game generation as the problem of producing a complete game artifact that realizes a specification through observable player-game interaction in a target environment. We argue that evaluating this setting requires three desiderata: Engine Grounding, Artifact Completeness, and Interactive Verification. We propose an interaction-grounded evaluation framework that assesses executable gameplay through replayed demonstrations and rubric-guided multimodal judging. We instantiate this framework as GameCraft-Bench, a benchmark comprising 140 Godot tasks across 15 game families. Evaluations of frontier coding agents show that end-to-end game generation remains highly challenging: the strongest agent achieves only 41.46%, and most agents score below 40%. Further analysis reveals that while agents often implement recognizable mechanics, they struggle to deliver complete games with sufficient content, functional visual feedback, and coherent presentation. See https://tongxuluo.github.io/gamecraft-bench-website for demos, code, and data.
Tongxu Luo, Rongsheng Wang, Jiaxi Bi +22
1The Chinese University of Hong Kong, Shenzhen · 2Shenzhen Loop Area Institute · 4USTB +5