Large language models have made automatic game generation increasingly practical, but taking a generated game beyond a merely playable state remains difficult. According to a new research paper, naive iterative refinement tends to overfit a small set of test cases, leaving fragile games with unresolved bugs, missing behaviors, and weak generalization to broader player interactions.
To address this, the authors introduce RSIGame, an autonomous agentic game development framework built around recursive self-improvement. RSIGame separates development into two complementary loops. A local explore-diagnose-improve loop explores the executable game broadly, diagnoses and prioritizes discovered issues, and performs evidence-grounded revisions, guided by an evolving checklist that accumulates new testing and improvement guidance over time.
A global loop tracks overall quality, preserves the best checkpoint, and detects saturation or regression across long-horizon development. Beyond test-time improvement, RSIGame also internalizes successful development experience into the generator through training.
The paper reports evaluations across 140 GameCraft-Bench tasks, two game engines, and five generators. RSIGame consistently improved game quality under matched development budgets. Notably, experience internalization enabled Qwen3.8-27B to reach 61.38 on Godot and 58.53 on Phaser, exceeding GPT-5.5 one-shot scores while reducing Qwen's generation tokens by 11 times.
The work was shared on Hugging Face alongside an interactive ResearchStudio Reel containing a visual poster, video, and blog in editable formats.