RSIGame: Autonomous Agentic Game Development with Recursive Self-improvement
RSIGame tries to keep generated games improving after they already run, without letting fixes collapse into narrow test-case wins.
The paper describes a two-loop agent framework: a local cycle that explores playable builds, diagnoses issues, and revises from evidence, plus a global cycle that tracks quality and guards against regression. Its checklist grows as development finds new failures and improvement targets. In tests across 140 GameCraft-Bench tasks, two engines, and five generators, the authors report consistent quality gains under matched budgets. They also say training on successful development experience let Qwen3.8-27B beat GPT-5.5 one-shot scores on Godot and Phaser while using far fewer generation tokens. HF Daily Papers' note
The paper describes a two-loop agent framework: a local cycle that explores playable builds, diagnoses issues, and revises from evidence, plus a global cycle that tracks quality and guards against regression. Its checklist grows as development finds new failures and improvement targets. In tests across 140 GameCraft-Bench tasks, two engines, and five generators, the authors report consistent quality gains under matched budgets. They also say training on successful development experience let Qwen3.8-27B beat GPT-5.5 one-shot scores on Godot and Phaser while using far fewer generation tokens. HF Daily Papers' note
score 4