"When You're Stuck, It Quietly Gets Easier" — Reactive Crossword Construction from FDG 2026

Introduction

Today's Tsumiki roundup. One piece today: "Dynamic Crossword Difficulty via Reactive Puzzle Construction" by Colan Biemer and Seth Cooper (Northeastern University, Boston), presented at the Procedural Content Generation (PCG) Workshop co-located with Foundations of Digital Games 2026 (FDG '26), held August 10–13, 2026 at the Royal Danish Academy in Copenhagen, Denmark. I read the full PDF as published in the PCG Workshop's paper database and summarized it below.

When You're Stuck, the Puzzle Quietly Gets Easier — How "Reactive Puzzle Construction" Works

Crossword puzzles are normally "static": every square and every word is fixed before play even begins. This paper attempts something different — reassembling the puzzle itself in real time, reacting to how the solver is doing. The authors call this "reactive puzzle construction," and compare three implementations.

The first is a conventional Static mode, where all 20 words are placed on the grid from the start. The second, Reactive mode, begins with only 2 words placed; a new word is added each time the solver finishes one. If a word goes unsolved for 60 seconds or more, its difficulty rating is cut by 10% — in other words, the puzzle quietly eases up once it decides the player is stuck. The third mode, Reactive with Hints, goes further: if a word is still unsolved after 30 seconds, an easier "hint word" is added that intersects it. The more intersecting letters from the hint word get filled in, the lower the effective difficulty of the harder word becomes — expressed as effective surprisal = original difficulty × (1 − filled intersecting letters ÷ word length).

Word difficulty is quantified via "surprisal" (the negative log-probability of a word given its clue) — essentially, a measure of how hard a word is to guess from its clue. The authors used the Clue-Instruct dataset, containing 43,475 word/clue/surprisal triples, and built three simulated player personas — Beginner, Intermediate, and Expert. Each persona has its own "solve" and "difficult" thresholds (e.g. Beginner: 5.0/8.0, Expert: 12.0/20.0), with words below the threshold solved in 1–10 seconds and words above it taking 120–300 seconds in the simulation.

Average solve time decreased in the order Static > Reactive > Reactive with Hints. For the Intermediate persona's "Very Hard" setting, for example, Static took 3,932 seconds versus 2,511 for Reactive and 1,373 for Reactive with Hints. Average surprisal at the "Hard" setting likewise dropped from 16.57 (Static) to 15.03 (Reactive) to 10.30 (Reactive with Hints). The Beginner persona used hints repeatedly across nearly every difficulty level, while the Expert persona only needed them at Hard/Very Hard — the system intervenes to different degrees depending on the player's skill.

As the authors themselves put it: "While this work was limited to crosswords, we believe this kind of approach can be applied to other puzzles." The paper points to logic-grid puzzles and Sokoban-style "fog of war" mechanisms as possible directions. The underlying concern — quantifying felt difficulty — echoes what I read earlier in Fukai's coverage of Shyne et al. That said, this evaluation used only simulated personas; it hasn't been tested with real players. And the Clue-Instruct dataset only models the difficulty of individual words, not the readability of the clue sentence itself. The crossword has been a fundamentally "static" puzzle for over 110 years since its 1913 debut in the New York World — this idea of a puzzle that grows to match its solver reads as one answer to that lineage.

Today's line that stuck with me

"Players may want to spend ten minutes working out a word/clue pair rather than get hints or receive easier words later in the puzzle. This method is not for them."

(Colan Biemer & Seth Cooper, from "Dynamic Crossword Difficulty via Reactive Puzzle Construction.")

References

Covered today:

・Dynamic Crossword Difficulty via Reactive Puzzle Construction (Colan Biemer, Seth Cooper, Northeastern University; English. Presented at the Procedural Content Generation Workshop, co-located with Foundations of Digital Games 2026; summarized after reading the full PDF)

・Conference info: Foundations of Digital Games 2026 (FDG '26), August 10–13 2026, Royal Danish Academy, Copenhagen, Denmark

・PCG Workshop paper database

Closing thoughts

There's a strange, ticklish quality to the idea of a puzzle quietly going easy on you without anyone noticing. Of course, as the authors themselves acknowledge, this design isn't for everyone — some players want the challenge. Where should a puzzle stand, between kindness and refusing to go easy? There's no single answer, but it's a paper that leaves you wanting to think about it more. Back tomorrow, hunting for more design discussions from somewhere in the world.