Researchers propose GameGo framework to improve coding-agent-based game generation

A new paper introduces GameGo, a framework designed to improve coding-agent-based game generation by systematically transforming sparse user queries into detailed product requirements documents grounded in industry practices. The authors argue that direct end-to-end game synthesis often leads coding agents to make underspecified assumptions, resulting in incomplete mechanics, disconnected gameplay flows, and limited visual aesthetics.

To address this, GameGo employs a scalable pipeline that converts brief game seeds into comprehensive requirements documents while using task-specific dynamic compression to retain core gameplay constraints without restricting design exploration. This method aims to maximize information density while preserving instruction-following fidelity during generation.

The researchers construct GameGoData, a dataset comprising 55,060 development trajectories spanning 2D, 2.5D, and 3D games, and introduce GameGoBench, a benchmark of 124 diverse game queries for evaluating coding agents. Training a model called GameGoCoder on GameGoData yields performance that surpasses matched baselines and is comparable to frontier models across gamedev benchmarks.

The authors state that all code, datasets, and models will be made publicly available, providing a foundation for further research and development in automated game creation.