TAG
#procedural-generation
2 reviews · 20 essays
Related essays
Zhou et al.: The Verifier is the Curriculum — Training Game Generation on a Launch Check Alone — Fukai Reads
A game-generation paper by Chenyu Zhou and colleagues. Starting from a diagnosis that the learned judge is gameable, they gate self-distillation on a single binary signal — does the generated Godot project launch cleanly — and over three rounds lift clean generation on four unseen families from 8.8% to 42.2%, with best-of-16 coverage going 18/25 to 25/25. Loosen the gate and the gain disappears.
Johnson et al.: What Changes in a Game When You Build an LLM Into It — Fukai Reads
A qualitative study by Johnson and colleagues at the University of Calgary on developing two games with an LLM embedded in their structure. Reading developer reflections, it analyzes how embedding an LLM as a component (not decoration) changes gameplay, playability, and player experience. Variability and personalization increase, but new burdens of correctness, difficulty calibration, and coherence emerge, with schema enforcement and validation as the keys.
Handcrafted or Generated — A Design Theory of Procedural Puzzle Levels
Who authors a puzzle's boards? I set the handcrafted lineage of Nikoli and Tametsi's 160 levels against the generated lineage of Simon Tatham's collection and Hexcells Infinite, ask what procedural generation drops, and look at the daily puzzle as a third way between generation and curation.
Making "solvable randomness": procedural content and the design of solvability in Google I/O 2026's Save the Date puzzle
One piece today. I read, in the original English, two official Google posts about this year's I/O Save the Date puzzle: "How we built the Google I/O 2026 Save the Date experience" (credited to Kacey Fahey and Caio Avelar, 3 March 2026) on the Google Developers Blog, and "How Googlers built the 2026 I/O save the date puzzle" (by Ari Marini, 6 March) on Google's The Keyword. The annual Save the Date puzzle — this year themed "Make Build Unlock" — is made of five cross-genre games plus a hidden sixth, Dino Pal. What interests me as a design matter is not the promotional shine but how the generated puzzles were kept solvable: Stretchy Cat reportedly uses "a level generation logic based on Hamiltonian pathing to produce random but solvable levels," Nonogram fixes level 1 while generating levels 2-3 on the fly, and Word Wheel generated 100 levels. That is the old, hard problem of generative puzzle design: random does not equal fun or fair. Unable to verify a fresh trusted source within the last few days, I cover this high-profile, first-party piece with its date (March) made explicit — while reading its design claims as a company's own account in what is, after all, a Gemini showcase.
Xu et al.: When Generative AI Becomes the Heart of Play — Fukai Reads the AI-Native Games Survey
A survey (arXiv preprint) by Zhiyue Xu and five co-authors on "AI-native games," where generative AI is the core loop itself. It defines them by a counterfactual — would play collapse if the AI were removed — and classifies 53 real artifacts along two axes: game type (G) and dominant AI mechanic (N), showing a skew toward narrative genres and a thin use of AI at the rule layer.
Letting an LLM build a whole game, and an AI playtest it: ScriptDoctor and the state of automatic game design
One piece today. I read, in the original English, "ScriptDoctor: Automatic Generation of PuzzleScript Games via Large Language Models and Tree Search" by Sam Earle, Julian Togelius and colleagues (arXiv:2506.06524; a short paper submitted to the IEEE Conference on Games). They pick PuzzleScript — the description language for turn-based 2D-grid puzzle games created by increpare (Stephen Lavelle) — as a "model organism," and have an LLM generate a whole game (rules, sprites, levels), iterating on it using compiler errors and the results of a breadth-first-search player agent. Feeding in a few human-authored games as examples clearly raises quality, and reasoning models (o1, o3-mini) beat GPT-4o. But the sharpest lesson is on the failure side: the games that looked most complex were often complex only because of broken mechanics — solvable is not the same as good. A rich read for anyone thinking about automatic game design.
"Solvable" and "legible": the two criteria for escape-room design that GenEscape spells out
One piece today. I read, in the original English, "GenEscape: Hierarchical Multi-Agent Generation of Escape Room Puzzles" by Mengyi Shan, Brian Curless, Ira Kemelmacher-Shlizerman and Steve Seitz of the University of Washington (arXiv:2506.21839). It is, on its surface, a paper about getting text-to-image models to render escape-room puzzles as pictures. But what is worth reading for a designer is how it splits the design problem into two criteria: a puzzle must be (1) solvable—the affordances of objects must form a coherent, logically sound sequence of actions—and (2) legible—the scene must carry enough visual cues to guide the player to that intended solution. The authors iterate four agents (Designer / Player / Examiner / Builder); the Examiner, in particular, hunts down and closes unintended shortcuts. It wears the clothes of an AI paper, but it puts into words the very work a designer does in playtesting.
Munk et al.: Generating Dynamic Game Text with Small Language Models — Fukai Reads
A paper by Munk et al. (IT University of Copenhagen) on generating in-game text dynamically with small language models (SLMs). It tackles the offline, cost and consistency walls of cloud LLMs using small models aggressively fine-tuned for narrow jobs. Their proof of concept, DefameLM, runs a medieval-RPG smear-poster loop, showing a one-billion-parameter-class model reaches high quality in a few seconds on a consumer PC.
Zeytuncu: Puzzle Difficulty Comes Down to How Many Numbers You Use — Fukai Reads
A difficulty-modeling paper by Yunus E. Zeytuncu on integer arithmetic puzzles (Countdown-style number games). Using an exact solver to generate over 3.4 million instances and defining difficulty by minimum operation count, it shows that the number of inputs used in a minimal solution alone is a 'minimal sufficient statistic' that perfectly determines difficulty.
Let the LLM Handle Story and Puzzles, Let the Symbolic Layer Keep the World From Breaking — Uruguay's IVIE on Incremental, Validated Generation of Interactive Fiction (ICCC'26)
One article today: IVIE, a paper headed to ICCC'26 by a team at the Universidad de la República in Uruguay (Vaucher, Silveira, Góngora, Chiruzzo), which I read in full in the original English on arXiv. The goal is to generate complete, playable interactive-fiction (text-adventure) worlds from scratch. The trick is a division of labor: creative decisions—setting, characters, puzzle design—go to an LLM, while a symbolic layer guarantees structural facts like spatial connectivity and objective solvability. Worlds are built backwards from the objective across four stages, each with a validation gate. In the puzzle stage, obstacle and solution are placed in different locations, solutions must be discoverable through exploration, and hints disclose in three escalating levels. Tellingly, in 3 of 16 evaluated worlds players slipped past puzzles simply by claiming they had solved them—surfacing a design tug-of-war: validate too strictly and you choke creativity; too loosely and the puzzles become hollow. Not a puzzle game per se, but a paper that touches the root of design: how to make 'validation' and 'freedom' coexist.
Where 'Solvable' and 'Fun' Diverge — PuzzleJAX Hands 500+ PuzzleScript Games to the Machines (arXiv, Aug 2025)
One article today: "PuzzleJAX: A Benchmark for Reasoning and Learning" (arXiv preprint, August 2025) by researchers at NYU, the University of Malta, the University of the Witwatersrand (South Africa), and Microsoft (Sam Earle, Graham Todd, Ahmed Khalifa, Julian Togelius and others). They reimplement PuzzleScript — Stephen Lavelle's (increpare) 2013 puzzle-authoring language — on the GPU and hand 500+ human-authored games to tree search, reinforcement learning, and large language models. Read as a designer, the core is one observation: 'solvable by a machine' and 'interesting to a human' are not the same thing. Tree search brute-forces simple games but stalls the moment they get richer; LLMs score 0% on most. The authors even note PuzzleScript's own creator hesitating to embed an auto-solver into the IDE, a caution about measuring difficulty by search.
Nasir et al.: Evolving the Rules of Play Themselves — Fukai Reads MORTAR
A paper on automatic game design by Nasir, Togelius and colleagues. Instead of levels, MORTAR evolves game mechanics themselves using a quality-diversity algorithm paired with a large language model, judging quality by whether stronger AI agents reliably beat weaker ones. Running on GPT-4o-mini, it generates diverse, playable games and even quantifies each mechanic's contribution.
Jiang et al.: Can a Sentence Build a Playable Game? — Fukai Reads OpenGame
A paper by Yilei Jiang et al. (CUHK) on OpenGame, an agent that generates whole 2D web games from natural language. Reusable skeletons and a 'living debug protocol' curb integration errors, setting a new state of the art across 150 tasks - though puzzles remained its weakest genre.
McConnell & Zhao: Generating Just-Right Puzzles in Real Time with a Genetic Algorithm — Fukai Reads
A paper by McConnell and Zhao on adaptive puzzle generation using a genetic algorithm. It generates Cosmic Express-style path puzzles in real time (about 7 seconds each) to match a player model built from how the player solves, and shows in an 18-person study that a time-only version lags on felt difficulty and sense of progression.
Li et al.: Can LLMs Play and Beat 2D Games? - Fukai Reads GVGAI-LLM
A paper by Li et al. (NYU and others) proposing GVGAI-LLM, a benchmark that has language models play 118 2D games to measure reasoning and spatial grounding. Translating boards into ASCII maps and solving zero-shot, GPT-4o-mini scored 0% on 477 of 540 levels and a 10.27% overall win rate, falling short of classic search algorithms. I unpack it as problem, method, findings, use cases, and limitations.
Kar: Using Autonomous Agents to Check at Runtime Whether Generated Levels Are Actually Playable — Fukai Reads
A PCG (procedural content generation) paper by Rishabh Kar of King's College London. It proposes Momentum, a mechanism that validates whether a generated course is actually traversable inside the same runtime loop, without pausing the game. Two autonomous agents run ahead of the player and inspect the path via geometric checks from the air and NavMesh checks on the ground. The evaluation is presented as structural estimates derived from the code.
Xu et al.: Promoting Game Mechanics to Coordinates to Generate Solvable Levels — Fukai Reads
A PCG (level generation) paper by Xu and Verbrugge of McGill University. Against geometry-first prior methods, it proposes HDPCG, which runs pathfinding on a dimensional-expanded graph that promotes mechanics such as gravity inversion and moving platforms to a coordinate, guaranteeing solvability during generation, and reproduces playable levels in Unity.
Feng et al.: Can AI Generate Counter-Intuitive Chess Puzzles? — Fukai Reads
A study, led by a Google DeepMind team, on generating creative chess puzzles with AI. A generative model trained on Lichess data is tuned with reinforcement learning, raising the rate of counter-intuitive puzzles from 0.22% to 2.5% (about tenfold). The highlight is how they reduce creativity to numbers a machine can measure.
Can AI Build a Whole Puzzle Game? ScriptDoctor and Its Generate-Playtest-Repair Loop
ScriptDoctor has a large language model write an entire puzzle game — rules, sprites, levels — then lets a compiler and a search-based agent inspect the result and demand revisions. The testbed is PuzzleScript, a language indie developers know well. I walk through the paper in five parts — problem, method, findings, where you can use it, limitations — covering why human-authored examples boost success rates, why reasoning models win, and the distance between 'solvable' and 'fun'.
How to make 'just-right' difficulty — letting a machine fit it to the player (a Canadian study) vs. a human authoring it through meaning (a US developer)
A version rebuilt with credible sources only. Two pieces today, both answering 'how do you deliver just-right difficulty?' from opposite directions. The first is a research paper by Canadian researchers Matthew McConnell and Richard Zhao (September 2025, arXiv): a system that generates puzzles in real time with a genetic algorithm and auto-tunes difficulty per player, validated in a user study. Its key finding: using 'time-on-task' alone as the adaptivity metric fails. The second is an interview with game designer Michael Hicks (Game Developer): churning out hard, time-consuming puzzles is easy; the truly hard part is finding interesting ideas to explore. A machine fitting difficulty to the player, and a human authoring difficulty through meaning. Both sources are peer-reviewed research and professional media - the kind makers can cite with confidence.
_studio.JPG)
Related reviews
Into the Breach
A top-down tactical puzzle: three mechs on an 8×8 grid, every enemy attack for the coming turn shown to you in advance, and a city to keep standing. Subset Games, makers of FTL, stripped almost all randomness out of the combat itself.
puzzlelogicgrid-basedShadows of Doubt
A first-person detective sim set in a procedurally generated 1980s noir city. Gather fingerprints, CCTV, receipts and call logs, pin them to an evidence board, and name the killer out of a fully simulated crowd — every citizen with a name, a job and a life of their own. From ColePowered Games.
mysteryimmersive-simfirst-person
