"Generate the Rules, Not the Levels" — RuleSweeper Has an AI Invent New Minesweeper Mechanics (IEEE CoG 2026)

Introduction

Today's Tsumiki roundup. One piece today.

IEEE Conference on Games (CoG) 2026, an academic conference on game AI research, ran from September 1 to 4, 2026, in Madrid, Spain. What caught my eye as a puzzle-design topic is a paper that tries to generate the rules themselves, not the boards.

Can an AI Invent New Ways to Play Minesweeper? — RuleSweeper (IEEE CoG 2026)

The short answer is: partially, yes. Minesweeper is best known as a logic puzzle where numbers tell you which cells are safe, but a paper presented at IEEE CoG 2026, called RuleSweeper, has an AI generate new rules for the game rather than new boards. After 100 generation cycles, it produced 51 rule variants that hold up as playable games in their own right. The authors are Ryan Fleishman, Shresth Kapoor, Teddy Clark, Jan Borowski, Tim Merino, and Julian Togelius, all at New York University. A write-up is available on the project site.

The starting point is this line from the authors: "Procedural content generation in games has traditionally focused on producing new levels within a fixed set of game mechanics, but there is much less work around generating the mechanics themselves." Most talk about automating puzzle design centers on cranking out more levels; this project instead hands the rule-design layer itself over to a machine.

Here's how it works. A large language model mutates a configuration object covering board size, mine count, how neighboring cells are counted, and win conditions, proposing new rule candidates. Each candidate is then handed to three kinds of players: a fully random agent, a traditional symbolic solver (PAFG), and an LLM-driven solver (PAFG-LLM). Only candidates where the gap between a competent solver and random play — what the paper calls "skill spread" — is at least 0.10 get archived as an actually playable game and become parents for the next round of mutation. Repeating this 100 times left 51 rules in the archive.

A few surviving rules make the intent clear: Radius + Drifting Mines, where clues are computed over a 5x5 area while mines shift position each turn; Telegraphed Mines, where some mines flash a warning in advance; Ranked Neighborhood, which reports relative rank instead of a raw mine count. Each keeps the core loop of counting numbers and avoiding mines, while changing exactly one thing about how information is given or how the board changes. Some variants reached a skill spread as high as 0.72, meaning these aren't just random noise.

On results, the LLM-driven solver (PAFG-LLM) beat the traditional solver's win rate on 40 of the 51 rules (78.4%). That suggests newly generated rules tend to resist solvers built around fixed procedures, and hints that rule generation itself could double as a design-testing tool.

Our own catalog already has 14 Minesweeper Variants, a game built from human-crafted Minesweeper rule variations. What RuleSweeper suggests is that machines might be able to take over the early, exploratory stage of that same process: finding and polishing a single rule. Treating the rule itself as the object of design is still a newer, thinner strand of discussion compared to level-generation research. It may point toward a division of labor where human designers focus on choosing which generated rule is worth developing further.

Today's Quote

"Procedural content generation in games has traditionally focused on producing new levels within a fixed set of game mechanics, but there is much less work around generating the mechanics themselves."

— from the RuleSweeper project write-up. In a field that mostly talks about generating more levels, this line stood out for pointing toward generating more rules instead.

Closing

I'm not good at solving puzzles myself, but as someone who dreams of designing them, the idea of generating the rules themselves genuinely excites me. If a machine can surface rule combinations no one thought of, that feels like a gift to the making side, not the solving side. See you again tomorrow.

Sources

Article covered today:

・RuleSweeper: Procedurally Generating Gameplay Mechanics In Minesweeper (Ryan Fleishman, Shresth Kapoor, Teddy Clark, Jan Borowski, Tim Merino, Julian Togelius, New York University; IEEE CoG 2026 project site)

・IEEE CoG 2026 Accepted Papers (accepted paper list; held September 1-4, 2026, Madrid)