RSIGame: AI Agents Self-Improve Game Development, 27B Model Beats GPT-5.5 One-Shot

RSIGame: Autonomous Agentic Game Development with Recursive Self-Improvement

Current large models can generate runnable games from natural language, but a runnable game is not a finished game. Real game development requires continuous playtesting, bug discovery, mechanic refinement, and content expansion. For AI agents, long-horizon iterative development is far harder than one-shot generation because game quality lacks simple pass/fail signals and agents must autonomously decide what to improve next and when to stop.

Dual-Loop Architecture: Local Loop and Global Loop

RSIGame addresses this with two coupled loops:

Local Loop – Autonomous issue discovery and per-edit verification. A Controller selects the next priority; an Explorer interacts with the game (keyboard/mouse) to uncover defects; findings enter a persistent development checklist; an Editor modifies code/assets; an independent Verifier re-interacts to confirm the fix actually improves behavior (e.g., attack button now damages enemies). This loop also explores new mechanics once basics work.

Global Loop – Long-term quality control via a Global Quality Monitor. It tracks quality across development phases, retains the best historical checkpoint to prevent regression, and detects saturation (diminishing returns) to halt or redirect development. This answers: where is improvement still valuable? Did this edit truly help? When should we pivot?

Experimental Results on GameCraft-Bench

Evaluated on 140 tasks across 15 game genres using Godot and Phaser engines with 5 initial generation settings . Quality measured by GameCraft-Bench's Overall score (mechanics, interaction, content depth).

RSIGame improves all base generators by 10.7–20.3 points over their raw one-shot output on Godot.

Under equal development budget, RSIGame outperforms iterative baseline Play2Code by 7.2–13.8 points .

Even starting from GPT-5.5 (50.26) , RSIGame reaches 64.53 , while plain iterative development plateaus early.

Quality curves show standard iterative development plateaus or degrades; RSIGame sustains gains. Global checkpoint retention and saturation detection are key.

Turning Development Process into Training Data

RSIGame collects a large-scale game-development experience corpus preserving the full trajectory from initial generation through multi-round edits:

2,213 game generation trajectories

2,108 planning trajectories

4,003 improvement rounds, of which 2,013 are independently verified as achieving their stated goal

This dataset captures how models plan, diagnose, edit, and verify – not just final code. Supervised fine-tuning of Qwen3.8-27B on Godot shows:

Base one-shot: 37.07

+ generation trajectories: 41.45

+ planning trajectories: 41.77

+ verified improvement experience: 48.22 (large gains in mechanics and gameplay depth)

Combined Effect: Experience-Trained 27B + RSIGame Surpasses GPT-5.5 One-Shot

Qwen3.8-27B (experience-trained) one-shot: 48.22 (near Codex+GPT-5.5's 50.26)

With RSIGame multi-round development: 61.38 on Godot, 58.53 on Phaser – both exceed GPT-5.5 one-shot (50.26 / 49.44).

Generation token cost drops from 6.41M to 0.57M (~1/11) on Godot.

Two-Layer Recursive Self-Improvement

Game Self-Improvement – Agent continuously refines the current game via exploration, editing, and verification.

Capability Self-Improvement – Verified development experience is fed back to train the base model, raising future one-shot quality and efficiency.

These loops reinforce each other: better models produce better starting games, which yield richer development experience, which further improves the model.

Qualitative Examples

Before/after GIFs demonstrate evolution across genres:

Fighting game – from bare combat prototype to complete versus experience.

Racing game – from basic driving scene to richer racing gameplay.

Shooter game – from simple shooting prototype to fuller combat levels.

Resources

Paper: https://arxiv.org/abs/2609.39045 Code: https://github.com/WenyiWU0111/RSIGame Project page / playable demos: https://huggingface.co/spaces/RSIGame/rsigame-page Hugging Face paper page:

https://huggingface.co/papers/2609.39045

Conclusion

RSIGame demonstrates that AI agents can move beyond one-shot generation to genuine long-horizon autonomous development. By structuring exploration, verification, and global quality control, agents sustainably improve game quality. Feeding verified development experience back into model training creates a second self-improvement loop, enabling smaller models (27B) to exceed larger models' one-shot performance. This points toward agents that continuously improve their own creations and accumulate capability from every verified iteration.