We (me, @smearle.bsky.social, @togelius.bsky.social) are working more to understand the relation between human solutions and the AI solutions in the #PuzzleScript games.

Help us by playing #online #puzzlescript games. No personal data is collected. Links are in the subsequent posts.

Work w/ Graham Todd, Yuchen Li, @amidos2006.bsky.social, Muhammad Umair Nasir, @zehuajiang.bsky.social, Andrzej Banburski-Fahey, and @togelius.bsky.social. Thanks to @increpare.bsky.social for creating and maintaining PuzzleScript, and to the many designers who have created beautiful things with it.

Much remains to be done. Can we lead LLMs to moments of creative discovery, allowing them to unlock trickier puzzles? Can we train robust RL players using curricula of levels/mechanics? Can we use feedback from diverse AI players to guide the synthesis of interesting new games?

Now, we can begin see how AI players respond to these challenges. The picture may be starkly surprising to some: simple tree search finds most solutions rapidly, while RL falls prey to obvious local minima, and LLMs spin their wheels when faced with unfamiliar semantics.

PuzzleScript games make for a great benchmark. Often, despite their mechanical simplicity, they elicit moments of insight in human players. Since 2013, casual and professional designers have brought considerable ingenuity to the language, generating a plethora of diverse games.

Paper: arxiv.org/abs/2508.16821
Code: github.com/smearle/scri...

PuzzleJAX is a faithful re-implementation of PuzzleScript (puzzlescript.net) capturing all of the engine's major features. It leverages the convolutional nature of rewrite rules to achieve major speedups in JAX.

We introduce PuzzleJAX, a benchmark for reasoning and learning. 🧩💡🦎

PuzzleJAX compiles hundreds of existing grid-based PuzzleScript games to hardware-accelerated JAX environments, and allows researchers to define new tasks via PuzzleScript's concise rewrite rule-based DSL.