RSIGame: Autonomous Agentic Game Development with Recursive Self-improvement
Wenyi Wu, Minghao Fu, Jieyu You, Kun Zhou, Siqi Liu, Aayush Salvi, Yiheng Lin, Ce Zhang and 5 others
University of California San Diego · ByteDance Inc. · Carnegie Mellon University
Published September 30, 2026 · on the feed October 1, 2026 · ▲ 83 · arXiv 2609.39045
The abstract
Recent advances in large language models have made automatic game generation increasingly feasible, yet reliably improving generated games beyond a playable version remains challenging. Naive iterative refinement can easily overfit a small set of test cases, producing fragile games with unresolved bugs, missing behaviors, and poor generalization to broader player interactions. We introduce RSIGame, an autonomous agentic game development framework with recursive self-improvement. RSIGame organizes development into complementary local and global loops. Concretely, a local explore-diagnose-improve loop broadly explores the executable game, diagnoses and prioritizes discovered issues, and performs evidence-grounded revision, where an evolving checklist continually accumulates new testing and improvement guidance. A global loop tracks overall quality, preserves the best checkpoint, and detects saturation or regression over long-horizon development. Beyond test-time improvement, RSIGame further internalizes successful development experience into the generator through training. Across 140 GameCraft-Bench tasks, two game engines, and five generators, RSIGame consistently improves game quality under matched development budgets. Notably, experience internalization enables 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 Brief
2:19A short podcast summary of the paper. Written and voiced by AI.
Transcript
RSIGame. Autonomous Agentic Game Development with Recursive Self-improvement. By Wenyi Wu and twelve others, from the University of California San Diego, ByteDance, and one other institution. Published on the thirtieth of September. Its code repository is worked out: about one hundred and thirty source files, with a readme and a licence.
Large language models have made automatic game generation increasingly feasible. Getting a game to playable is one thing. Reliably improving it past that point is another, and it remains hard. Naive iterative refinement can overfit a small set of test cases. What comes out is fragile: bugs left unresolved, behaviors missing, and poor generalization once players interact with the game in broader ways.
RSIGame is an autonomous agentic framework for game development built around recursive self-improvement. It organizes development into two loops that complement each other, one local and one global, and then goes a step further by feeding what worked back into the generator itself.
The local loop runs explore, diagnose, improve. It explores the executable game broadly, diagnoses the issues it finds and prioritizes them, and makes revisions grounded in evidence. Throughout, an evolving checklist keeps accumulating new guidance for testing and for improvement. The global loop watches the larger arc. It tracks overall quality, keeps the best checkpoint, and detects when long-horizon development has saturated or started to regress. Beyond improvement at test time, RSIGame also internalizes successful development experience into the generator through training.
Evaluation spans one hundred and forty GameCraft-Bench tasks, two game engines, and five generators. Under matched development budgets, RSIGame consistently improves game quality. The internalization step stands out. With it, Qwen3.8-27B reaches sixty-one point three eight on Godot and fifty-eight point five three on Phaser. Both exceed the one-shot scores of GPT-5.5, and Qwen's generation tokens drop by a factor of eleven.
Two kinds of memory run through this design. The checklist and the saved checkpoint carry lessons within a single development run. Internalization moves them into the generator itself through training, and in the reported numbers that is the step that lets Qwen outscore the one-shot attempt of GPT-5.5 while cutting its own generation tokens by a factor of eleven.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Playable is easy now. Good is the wall.
You have a language model that can turn a game description into a playable game. Getting past playable to good is the hard part.
The road ahead
- 1 · The problem ~7 min A language model can already make a playable game, but a simple test-and-fix loop stops at passing a few tests, not at a good game.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Test widely by exploring the running game, fix what the evidence shows, record what you learn in a growing checklist, and let an outer loop guard overall quality.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Explore the game, diagnose and rank issues, edit from evidence, keep the best checkpoint, then train the generator on what worked.
- 6 · The experiment ~6 min You’ll guess: After experience internalization, by what factor did Qwen's generation tokens drop?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.