RSIGame tackles the challenge of reliably improving automatically generated games beyond a basic playable version. Naive iterative refinement often overfits to a small set of test cases, leaving bugs and missing behaviors. RSIGame uses a local explore-diagnose-improve loop that broadly tests the game, diagnoses issues, and updates a checklist, paired with a global loop that tracks overall quality and prevents regression. Successful development experiences are also internalized into the generator via training.
Tested on 140 GameCraft-Bench tasks across two engines and five generators, RSIGame consistently improved game quality within a fixed development budget. Notably, the Qwen3.8-27B model using experience internalization achieved scores of 61.38 on Godot and 58.53 on Phaser, surpassing GPT-5.5 one-shot scores while using 11 times fewer generation tokens.