RSIGame

RSIGame is a framework for autonomous game development that employs a recursive self-improvement approach to enable large models to autonomously develop games.

RSIGame organizes development into complementary local loops and global loops.

RSIGame's local loop performs fine-grained recursive improvement through three phases—autonomous exploration, problem diagnosis, and improvement—where the autonomous exploration phase involves collaboration between two agents: a controller and an explorer.

RSIGame's global loop includes a game quality monitor and a convergence control mechanism, with the quality monitor fully isolated from the benchmark evaluator to ensure impartiality.

The authors obtained 2,213 generation trajectories, 2,108 planning trajectories, and 2,013 rounds of validation improvement data from 4,003 candidates, which were used for supervised fine-tuning of Qwen3-8B-27B.

On the Godot engine, RSIGame achieved an overall score improvement of 10.7 to 20.3 points over the frozen initial game across all five generator configurations.

Under the same computational budget, RSIGame outperforms the comparison method Play2Code by 7.2 to 13.8 points.

Experience internalization fine-tuning improved Qwen3-8B-27B's one-shot score from 37.07 to 48.22, while reducing generated tokens from 6.41 million to 570,000 (an 11-fold reduction).

With RSIGame applied on top, the score rose to 61.38, surpassing Codex+GPT 5.5's one-shot generation score of 50.26, with total token usage reduced by 2.6 times.

Adaptive evolution outperforms fixed scheduling, with effort reallocation based on current game state receiving 59% preference in pairwise evaluation, compared to only 37% for fixed round-robin scheduling.

Evidence-driven pre-improvement validation improved precision from 58.6% to 72.3%, and reduced unsupported objectives per round from 1.93 to 0.50.

Replay-based validation can detect 76.2% of unsuccessful improvements, whereas build-check-only approaches fail to detect any of these behavioral failures.