I fine-tuned GPT-2-XL with LoRA to generate playable levels for my Bloxorz-inspired puzzle game (Mindcraft).
Based on the "Level generation through large language models" paper (NYU, 2023) which did this for Sokoban. I adapted their approach to work with block-rolling puzzles.
The interesting part: I didn't give it any solution data during training - just level layouts and metadata (grid size, move count, gimmick types). After 10k steps, it generated 22% valid+novel levels. With 50k steps on levels with glass tiles, that jumped to 64%.
The model learns what makes a level solvable just from seeing enough examples. It's not perfect (grid size accuracy is low), but the generated levels work in the actual game.
Trained on RTX 4080 (16GB) using LoRA to keep it feasible on consumer hardware.
I teach students at a university. Recently, in my JavaScript class, I wanted to explain mouse and keyboard input in a visual way, but I couldn't find a suitable example.
So I created one myself, and I'm sharing it here in case it might be helpful to you as well.
Please test it out and let me know if you have any feedback.
Thank you for your interest in my previous post. This time, I've written a blog post about the game and the process of creating it.
In the original Minesweeper, there are inevitable 50/50 moments where you have to rely on luck. In the game I created, 'Explainable Minesweeper,' I eliminated these guessing situations. However, I also prevented the maps from becoming too easy! How? By using logical deduction, you can solve puzzles that initially appear to be luck-based. The blog post explains the process in more detail.
I hit an interesting problem with my puzzle game Logic Islands - 3 out of 6 rulesets would hang forever trying to generate maps larger than 7x7.
The trick that worked? Using Wave Function Collapse, but choosing what to generate based on each ruleset - islands for some, walls for others. This flexibility made complex constraints (like "no 2x2 blocks") trivial to express as tile connection rules.
My favorite result: the "Minimal" ruleset enforces "all wall regions must be exactly 3 cells" using just 11 tiles and local WFC constraints. No post-processing needed.
Now generates 12x12 maps instantly instead of hanging forever.
Anyone else using WFC for logic puzzles beyond typical texture synthesis?
I wonder how well it will work for generating certain street tile patterns, where tiles of different sizes are used and where it is not allowed to have four tiles meeting at one point and where there are no H-patterns. See [1] for a large pattern and [2] for an animation using patterns within an 8 by 8 square. I did figure out a set of Wang tiles [3].
Nice to meet you. It seems that you have been researching this topic in depth. Since you have been researching this topic for a long time, I don't have any immediate thoughts on it, but I think I need to think about it a little more.
While working on Simple-Tiled WFC this time, I kept wondering whether I should reference neighbours in more than four directions, but in the end, I'm glad I finished without referencing them. I hope this Random Street Tile Pattern can also be solved in such an elegant way!
I analyzed 7 years of Armorgames.com data (999 games) to understand web gaming market trends.
Key findings that might interest fellow developers:
User standards are rising: Average ratings dropped from 7.02 (2018) to 6.45 (2025), but the percentage of high-quality games (8.5+ rating) actually increased from 12.3% to 14.7%. This suggests quality polarization rather than overall decline.
Genre trends:
Rising: Idle games, Strategy, RPGs (deeper gameplay mechanics)
Declining: Traditional arcade/action games
Stable: Puzzle and Adventure (web gaming staples)
Innovation wins: The highest-rated "hidden gems" all had one thing in common - innovative mechanics rather than genre variations. Games like "Detective Bass: Fish Out of Water" (9.3 rating) and "SYNTAXIA" (9.1 rating) show originality still pays off.
Market maturation: The correlation between rating and popularity is surprisingly weak (0.126), suggesting quality ≠ virality. However, play count strongly correlates with favorites (0.712).
I'm not sure the take away for the first point that user standards are rising is correct. Could that also be the number of people making games is increasing? I say this because more highly rated games and a trending down of the average (more slop) could explain that as well. I think the idea that standards are rising would hold constant the number of games.
Maybe on other platforms? Armorgames curates pretty heavily, but you're right that AI-generated games could be flooding less selective platforms. Would be interesting to run this same analysis on Steam or itch.io where the barriers are lower.
Fair point, but here's the thing - Armorgames is actually way pickier now about which games they accept. They're letting fewer games through their gates. So if the average rating is still dropping even with higher curation standards, that pretty much confirms users have gotten more critical over time.
Based on the "Level generation through large language models" paper (NYU, 2023) which did this for Sokoban. I adapted their approach to work with block-rolling puzzles.
The interesting part: I didn't give it any solution data during training - just level layouts and metadata (grid size, move count, gimmick types). After 10k steps, it generated 22% valid+novel levels. With 50k steps on levels with glass tiles, that jumped to 64%.
The model learns what makes a level solvable just from seeing enough examples. It's not perfect (grid size accuracy is low), but the generated levels work in the actual game.
Trained on RTX 4080 (16GB) using LoRA to keep it feasible on consumer hardware.