Has AI made it practical to automatically score the quality of puzzle designs?
retained text
I'm currently developing a puzzle game, and I've been thinking about procedural puzzle generation. One problem I keep coming back to is: how do you automatically tell whether a generated puzzle is actually good? I looked around a bit, but a lot of the discussions and research I found seemed relatively old. That made me wonder whether recent AI/LLM-based approaches have changed what's practical here. Traditionally, I imagine you could score things that are relatively easy to measure: number of solutions, solution length, redundant moves, branching factor, required mechanics, and so on. But could AI now help evaluate some of the harder-to-quantify aspects of puzzle design as well? Things like whether the solution requires an interesting deduction, whether there's an "aha" moment, whether mechanics interact in meaningful ways, or whether a puzzle is technically difficult but just tedious. If we could combine those evaluations into a reasonably reliable "puzzle quality score," it seems like procedural generation could become much more practical: generate thousands of puzzles, automatically evaluate them, and only keep the promising ones for human review. Has anyone experimented with something like this recently?
document fetched 2026-09-24 00:28 · canonical https://reddit.com/r/gamedev/comments/1wo5s6k/has_ai_made_it_practical_to_automatically_score