Coding agents build and repair games with limited accuracy
SWE-Game: Can Coding Agents Build the Games We Want?
Artificial Intelligence
Summary
Making video games is hard and requires careful coding to work right. The authors created SWE-Game, a set of tasks and tests where computer programs try to make or fix games following specific instructions. They tested different AI models and found that while some can build playable games, their accuracy is still below 60 out of 100, with many mistakes in following requirements and gameplay logic. The study also developed ways to check if the games work as intended by replaying inputs and using visual checks that match human judgments fairly well.
What this means in practice
- •For game developers: Evaluate and improve AI tools that write or fix game code automatically across different 2D and 3D game types.
- •For software testing teams: Use executable game checks combined with visual assessments to verify game mechanics and repairs from AI-generated code.
Authors
Xiaoyu Chen, Lai Wei, Jin Wang, Xiangyu Zou, Ruochen Fan, Enze Luo, Mingzhe Yao, Jiahui Zhu, Yuhua Wen, Linghe Kong, Weiran Huang
Abstract
We introduce SWE-Game, a benchmark of 247 tasks grounded in 41 executable reference Godot games spanning 13 gameplay categories in 2D and 3D. Five task types cover development from a brief, implementation from a game design document, skeleton completion, repair of 83 injected-fault cases, and Godot-to-Unity porting. Reference materials specify the intended gameplay, while a shared instrumentation interface lets evaluator-owned drivers and probes execute actions and observe independently implemented games. Evaluation combines engine-state checks, certified reference-input replay, and agent-authored feature demonstrations to assess mechanic correctness, demonstrated playability, and behavioral restoration and preservation after repairs. Game-specific vision-language rubrics separately assess presentation. Across six models, Opus5 achieves the highest overall score in all five task types. Best overall scores remain below 60 out of 100 across the three construction tasks, with Brief-to-Game reaching 50.38. Analysis of reviewed submissions identifies requirement omissions and gameplay logic errors as predominant implementation problems. On human-labeled behaviors from 100 agent-built games, executable checks achieve 92.59% balanced accuracy, compared with 78.41% for a video-based VLM judge. Rubric-based visual scores reach a Spearman correlation of 0.829 with human ratings of 200 gameplay clips. Together, these results characterize current agent capabilities across game-development activities and support combining runtime evidence with visual assessment.