Chao et al.: Insight Is About Searching Far — Fukai Reads

TL;DR

Where does insight come from? Insight is the sudden "I've got it!" understanding of a problem, and its felt quality is the "Aha" experience. This paper uses a Japanese version of the Remote Associates Test (RAT - find a fourth kanji that joins three seemingly unrelated kanji), with 349 and 105 participants plus a computer simulation, to numerically track the search path that leads to a solution.

The central finding: "de-fixation" - setting aside a wrong idea so you can pursue alternatives - is necessary for solving, but it is not what decides whether insight occurs. The hallmark of insight was exploring the solution space over greater distances, i.e. holding a wider set of candidate solutions for evaluation. The harder the problem, the larger the optimal breadth of exploration. This article aims to convey the key points without your having to open the paper.

Introduction

Today's paper is "Long-distance exploration in insightful problem-solving" by Zenas C. Chao, Feng-Yang Hsieh, and Chien-Te Wu. It appeared in Communications Psychology, a peer-reviewed Nature Portfolio journal, volume 3, article 53 (2025), published 26 March 2025, open access. According to the article metadata, co-author Wu is affiliated with the University of Florida's brain institute (Center for Cognitive Aging and Memory, McKnight Brain Institute). This is a peer-reviewed paper, not a preprint.

I usually scan the new AI listings on arXiv, but today I deliberately picked a cognitive-psychology paper. Two reasons. First, the concepts it studies - insight, fixation, exploration - are exactly the vocabulary of puzzle design. Second, for someone who makes puzzles, understanding what happens inside the human head is as useful as, if not more useful than, automatic content generation. I brewed a strong cup of hot drip coffee and read it with a pen on a printed PDF.

For honesty: more than a year has passed since publication (this series usually favors very recent papers). But foundational findings in cognitive science age slowly, and citations are steadily accruing (the article page shows 6,028 accesses and 5 citations). I judged its takeaways for design to be valuable enough to feature.

Background

Insight is described as: an impasse, then a restructuring of how the problem is seen, then a sudden appearance of the solution. By contrast, a step-by-step "analytic" approach yields a weaker Aha. So how does the in-the-head search proceed? Two theories have long competed. Constraint-relaxation theory holds that dissolving the assumptions (constraints) that block the solution opens new regions - the de-fixation story. Progress-monitoring theory holds that you continually check whether the search is making progress and switch strategy when stuck - closer to a story about breadth of exploration.

What was unknown is how de-fixation and exploration actually interweave during insight. Creative problem-solving is complex, and quantifying the mind's movements from the outside is hard. So the authors turn to the RAT (more precisely a compound remote associates, CRA, task: find the answer common to three words). The RAT has a single defined answer, and the strength of each word's link to the answer can be measured from a language corpus (a co-occurrence dictionary built from huge amounts of text), making it one of the few tasks where you can numerically follow how the search unfolds.

Why does this matter for puzzle design? Much of what we call a "good puzzle" is precisely engineered around this impasse and the momentary restructuring that breaks it. How to build fixation, where to let it dissolve, where to trigger the Aha - the raw materials for those questions arrive here in numerical form.

Approach

The authors combine two experiments with a simulation. Experiment 1 is a "fixation-controlled RAT (FC-RAT)." Alongside the three question kanji, a Fixation condition shows "fixation cues" that each form a two-kanji compound with a question kanji, compared against a Neutral condition showing unrelated kanji. Crucially, the cues are not the answer - they are placed as obstacles to it (in English: if the question is Board / Magic / Death with answer Black, the cues might be Game / Show / Match).

Difficulty, put in words rather than equations: the paper measures how strongly each question word is linked to the answer using corpus co-occurrence frequency, treating weaker (more distant) links as harder (the "distance to answer"). It similarly measures how close the cues are to the question words (the "distance to cues"). Experiment 2 is a "thought-tracing RAT (TT-RAT)," where participants type each kanji that comes to mind one at a time, and the average distance of those thoughts from the question words is computed as the "exploration distance."

Third, they build a simulation model that moves around a corpus of words to solve RATs. It has three knobs - how easily one gets fixated, de-fixation (the power to suppress a wrong thought), and exploration capacity (how many options can be held as candidates at once). The authors checked in advance, via simulation, whether they had enough trials to detect effects (a power analysis), and report Bayes factors and effect sizes (Cohen's d, the size of a difference in standardized units). They also state explicitly that the study was not preregistered (publicly registering hypotheses and analysis plan before running it).

Findings

In Experiment 1, the fixation cues did bite. Accuracy fell to 39.5% in the Fixation condition versus 53.8% in the Neutral condition (a fairly large effect, Cohen's d = -0.80), and reaction time grew from 14.7 s (Neutral) to 18.5 s (Fixation). The rate of failing to answer in time was also higher under fixation (33.1% vs 21.8%). And yet - this is the crux - the rate of Aha experiences was 59.8% (Fixation) versus 61.8% (Neutral), with no statistical difference (the authors report Bayes factor BF10 = 0.10 as evidence of no difference). The cues made solving harder and slower, but did not change whether a solution was reached by insight.

In Experiment 2, among correctly solved trials, the exploration distance was 5.7 for trials with an Aha versus 5.4 without, significantly larger for Aha (d = 0.24, BF10 = 34.22). Meanwhile, the number of thoughts before reaching the answer did not differ by Aha. From this the authors organize the idea that insight ties to the distance of exploration, not the quantity of thoughts (see the paper's Figure 6).

In the simulation (1,975,000 runs total), reaction time grew with stronger fixation and shrank with stronger de-fixation and greater exploration capacity. Moreover, each problem had an "optimal exploration capacity" that minimized reaction time, and it correlated strongly with problem difficulty (r = 0.637): the harder the problem, the more broadly you should explore to solve it faster. The model also reproduced the experiments' surprising results - that nearer cues actually sped up insightful solving, and that the relationship between cue distance and accuracy flips depending on the time limit. Note that the insight-specific effects (such as the exploration-distance difference) are on the small side in effect-size terms (d ~ 0.24, correlation r ~ 0.14), and I report them at the paper's own numbers.

Use cases

If I were building a word puzzle (associations, anagrams, chain-of-words), I'd first repurpose the "distance to answer" idea for difficulty estimation: measure the corpus co-occurrence between words and the answer, and treat weaker links as harder - this paper backs that relationship with real data. For a daily word game, simply sorting candidate prompts by this distance gives you a difficulty curve.

If I were building a hyper-casual riddle or "trick" puzzle, the design of deliberately placed fixation cues - near-miss traps - gains support. In the paper, nearer cues made solving slower and harder but did not lower the chance of insight, and actually sped up insightful searches. Plausible decoys are thus a way to add difficulty without cutting the satisfaction (Aha). But watch the interaction with the time limit: the data show a flip where, under a short limit, "near traps" reduce accuracy, while under a long limit, "far traps" do. Timer length and trap placement should be designed together.

If I could log player behavior, I'd record not the number of attempts but the distance of attempts (how semantically or spatially far apart the moves are) as an insight signal. In the paper, the number of thoughts was unrelated to Aha; only the exploration distance mattered. As a design heuristic: since harder levels have a wider optimal exploration breadth (r = 0.637), at the hard spots it may help to give players a state where they can survey broadly rather than narrowing their options and tools, which reads as favoring faster solving and insight. Whether this transfers directly to spatial puzzles like Sokoban-likes is a separate question (below), but it's well worth testing as a design hypothesis.

Limitations

Start with what the authors themselves acknowledge. They state the study was not preregistered. They also use the Aha experience as an indicator of insight while noting it does not necessarily mark "true" insight - an Aha can accompany a wrong answer (a so-called false insight), and an Aha can bias judgments of a solution's quality. Further, the model is built to reproduce average group behavior rather than individual differences, and the authors are explicit that it simplifies real thought - e.g. delivering all question words at once, and representing fixation as a constant amplification.

What I (Fukai) would flag: first, the small effect sizes. The fixation effect is large (d = -0.80), but the headline "insight = exploration distance" effect is modest (d ~ 0.24, correlation r ~ 0.14). With large samples and strong Bayes factors, that the effect exists is trustworthy; how strong it is should be received cautiously. Second, the task is limited to Japanese compound remote associates (CRA). Whether the same law holds for spatial or mechanical puzzles like Sokoban, or for narrative riddles, cannot be settled from this paper alone. Finally, what is shown is correlation, not the causal claim that "widening exploration increases insight" - the reverse order (being in an insight-prone state allowed broader exploration) remains possible.

Fukai's Reading

From here, I flag this as my own interpretation. I would place this study in the lineage that frames creativity as "search within a solution space" - the line that runs through small-world network models and the like. In the vocabulary of design criticism, the paper reads as giving numerical backing to the claim that "insight is a property not of the answer itself but of the trajectory that reaches it." If so, what a puzzle author designs is not only the distance to the answer. How wide a leap the player is placed in a position to make - that "room to leap" is, I take it, the very thing we design when we design what we call the Aha.

Closing

If you want to go deeper, reading a recent review that organizes the "restructuring" of insight (a 2023 review on the Aha experience in Nature Reviews Psychology) alongside Danek & Wiley (2017), which dissects the "false insight" where an Aha does not mean a correct answer, will help the map come into view. For the foundations of the RAT itself, Bowden & Jung-Beeman's normative data for compound remote associates (2003) is a starting point. Next time too, I'll pick a paper that someone making games can read and put to concrete use.

References

Papers and related materials referenced in this article:

・Long-distance exploration in insightful problem-solving (Chao, Hsieh & Wu, 2025, Communications Psychology vol.3, art.53, peer-reviewed)

・DOI: 10.1038/s44271-025-00235-4

・Related: Restructuring processes and Aha! experiences in insight problem solving (Nature Reviews Psychology, 2023)

・Related: Danek & Wiley, What about false insights? (Frontiers in Psychology 7:2077, 2017) / Bowden & Jung-Beeman, Normative data for 144 compound remote associate problems (Behavior Research Methods, 2003)