AI Papers Reader

Personalized digests of latest AI research

View on GitHub

VibeGame Explores Turning Text Prompts Into Playable 2D Games

Building a video game has traditionally taken months of coordinated work from designers, programmers, and artists. AI coding agents have made ordinary software easier to produce from plain-language requests, but games remain harder because code, scenes, art, animation, physics, and settings must all stay consistent, and many flaws appear only during play. Researchers from Nanjing University, Nanyang Technological University, and Liblib.ai report a system called VibeGame that builds complete 2D game projects from text prompts and optional reference images.

The central idea is to give AI agents a game they can inspect and test while they work. VibeGame’s engine, built on the open-source Phaser web framework, stores each project as structured text, including scenes, nodes (units that combine visuals, collision shapes, scripts, and settings), assets, and configuration files. A schema check catches problems such as missing assets before the game runs. The engine also lets an agent pause the game, advance it frame by frame, inject actions such as “press jump,” and capture screenshots. The authors argue that this turns real-time play into a process an agent can verify step by step.

Built on this engine is what the authors call an Adversarial Agent Team, eight specialized roles that separate building from judging. A designer expands a request into a design document, an artist generates art, and programmers implement the code. An auditor checks the work, a player tests the running game, and a reviewer acts as the final gate. Finally, a training-free “self-evolution” step distills accepted projects into reusable templates, modules, and workflow contracts for future projects.

The paper offers one illustrative case. In a boxing task, the opponent must telegraph each punch so the player can react. The Codex agent working in the prompt-assets-to-game (PA2G) setting, which supplies an asset library, drew its boxers with programmatic shapes and rendered punches as stretched rectangles. The Codex agent in the prompt-to-game (P2G) setting produced polished visuals but showed no wind-up. VibeGame’s version included visible wind-up tells.

Testing used 30 tasks from the GameCraft-Bench benchmark, two from each of 15 game types, with a 24-hour budget per task. The judge was GPT-6.1-Sol, which the authors chose after finding judgment errors in GPT-5.5. With the GPT-5.6-Luna model, VibeGame scored 49.99 overall, 11.70 points above the strongest baseline, Qwen Code in P2G mode (38.29). With the frontier model GPT-6.1-Sol, it scored 72.99, 4.19 points above Codex in P2G mode (68.80). Most of that gain came from content depth (+4.76 points) and art and presentation (+7.72 points). VibeGame trailed slightly on core mechanics (86.26 versus 87.18) and functional visuals (62.29 versus 62.48). In a blind study with 30 participants, human evaluators preferred VibeGame in an average of 72.7% of overall comparisons.

An ablation on 10 tasks using GPT-5.6-Luna suggests both the engine and the review roles matter. Removing the engine lowered the score from 53.45 to 43.12, and removing the adversarial review lowered it to 48.82. Adding experience distilled by GPT-6.1-Sol raised the score to 56.37.

The study has clear limits. It covers only 2D games built on a web engine, not 3D or long-lived commercial titles. VibeGame outperformed the strongest baseline in 13 of 15 game types, and the paper does not report token or dollar costs for its long run times. The results also rest on one benchmark subset and a model-based judge, although human preferences broadly agreed.

Still, the work suggests that how an agent can see and test a project may matter as much as the underlying model. For now, that remains the authors’ interpretation of their experiments.