On GameCraft-Bench's Godot tasks, RSIGame improves games initialized with Codex and GPT-5.5 from an Overall quality score of 50.26 to 64.53 through autonomous development.

Researchers from UC San Diego, ByteDance and Carnegie Mellon University propose two cooperating loops to make iteration useful. The goal is to improve playable games through observed behavior, rather than repeatedly rewriting code without checking the result.

A local loop explores the game, diagnoses problems, edits code and verifies fixes through interaction. A shared checklist tracks evidence and regressions. A global monitor compares progress, preserves the best checkpoint and detects when a development direction has stopped paying off.

The team also uses verified development experience to train Qwen3.8-27B. Combined with RSIGame, it reaches 61.38 on Godot, above the 50.26 one-shot GPT-5.5 baseline. This compares different systems and development budgets, not equally resourced base models. The scores measure benchmark game quality, not universal coding ability or human player satisfaction.

RSIGame: Autonomous Agentic Game Development with Recursive Self-improvement

Paper: https://arxiv.org/abs/2609.39045
Project: https://huggingface.co/spaces/RSIGame/rsigame-page
Code: https://github.com/WenyiWU0111/RSIGame
Our Report: https://www.jiqizhixin.com/articles/2026-10-21

📬 #PapersAccepted by Jiqizhixin