VibeGame tops general LLM harnesses on 2D game benchmark

A purpose-built framework beats general-purpose agent harnesses using the same underlying language model at building complete 2D games from prompts, the authors of VibeGame report.

The design has two unusual pieces. The game engine itself represents project state as "schema-validated structured text" and provides "frame-synchronous runtime control," which the paper says lets agents "inspect game states, execute actions, and verify behavior at their own pace." Around that engine sits what the authors call an "Adversarial Agent Team of eight specialized agents to separate implementation from review, so that problems are found and fixed through multi-round loops."

A third component is a "training-free self-evolution mechanism" that distills accepted projects into "reusable skeletons, modules, and collaboration contracts for future development."

The headline result, in the paper's own words: "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." P2G is the harder of two settings, in which the agents also have to generate the art rather than drawing from a supplied asset library. The abstract names no raw scores and does not state the size of the GameCraft-Bench subset used.