Nasvytis & Fan: Insight and Transfer Show Up in How You Talk — Fukai Reads

TL;DR

Sometimes a problem that felt hopeless suddenly gives way — an 'insight' (the moment the way you see a problem flips and the answer appears). This paper studies such insights and how a knack discovered in one problem carries over to similar problems (transfer), using an experiment where people think aloud while solving.

The authors had 189 participants solve five matchstick-arithmetic puzzles in a row, automatically transcribing and analyzing what they said. When the same problem type recurred, people got faster and more accurate after their first success, and increasingly said out loud 'this is the same type as before.'

In other words, when transfer is happening, people become able to put the knack into words. I read this as an experimental backing for the level-design maxim of teaching one mechanic by repetition before moving on.

Introduction

The authors are Linas Nasvytis and Judith E. Fan, both at Stanford University (Psychology, the Graduate School of Education, and Computer Science). This is a preprint posted to arXiv in May 2026 (arXiv:2605.12970) — a pre-review manuscript in the format of a short cognitive-science paper — and I cannot confirm it has passed peer review. That said, the study is preregistered (hypotheses and analysis plan posted before running, to prevent after-the-fact analysis), and its data and code are public, which I appreciate.

I chose it today because the subject is puzzles themselves. The task is matchstick arithmetic, a classic insight problem, and its material speaks directly to puzzle makers. The finding that 'repeating the same type speeds mastery' bears directly on how you order levels. After a run of AI papers, today I wanted one that looks at human cognition itself.

Background

Insight has long been described as a discontinuous leap accompanied by a sudden 'Aha.' An influential account in cognitive science holds that insight is a switch of the search arena from an initial way of framing the problem (the problem representation) to another (Kaplan & Simon, 1990) — brought about by relaxing implicit assumptions or breaking familiar chunks apart (Knoblich et al., 1999).

The trouble is that this switch happens in the head and cannot be observed directly. Researchers have used the think-aloud protocol (having people voice their thoughts as they work), but transcribing and coding recordings was laborious and hard to scale.

Recent advances in automatic speech recognition and natural language processing (NLP) change that. It is now feasible, at scale, to analyze not only how much and how long people talk but what they talk about and what role each utterance plays in reasoning. This paper puts those tools to work on insight and transfer.

Approach

The task is matchstick arithmetic: an invalid Roman-numeral equation drawn in matchsticks that becomes valid by moving a single stick. The authors created 25 new problems across five transformation types. Participants were 202 people recruited via crowdsourcing (Prolific); after preregistered exclusions (failed practice, bad audio, missing data), 189 were analyzed.

The crux is the grouping. The 'Same' group saw all five problems of one transformation type, so the type repeated. The 'Different' group saw a different type each time, so it never repeated. This separates two stages: the first success (which the authors treat as a marker that a re-framing occurred) and whether the knack could be reused afterward. Each problem had a four-minute limit, and participants thought aloud throughout while being recorded.

The analysis is multi-layered. Silero voice-activity detection measures how much they spoke (speech density). WhisperX transcribes; each utterance is turned into numbers by OpenAI's embedding model (text-embedding-3-large, which converts a sentence's meaning into a sequence of numbers), and classifiers test whether correctness or phase can be predicted. GPT-5.1 then labels each utterance with one of seven reasoning moves (proposal, evaluation, categorization, restatement, affect, filler, other). No equations are used; the design reads changes in how people talk.

Findings

First, the difficulty range was confirmed. Overall accuracy was 47.5% (449 of 945 trials). By type, 'value-loose-chunk' was easiest at 80.7% and 'tautological' hardest at 19.2% — a spread the authors call well-suited to studying learning and transfer.

The central result: in the Same group, accuracy rose from 0.32 on trial 1 to 0.75 on trial 5, while the Different group stayed between 0.32 and 0.39 (trial-by-group interaction beta=-0.524, p<.001). Among those who first solved by trial 4, the Same group then solved 83.7% of later problems versus 44.0% for Different, and did so faster (89.9 s vs 179.0 s).

Speech changed too. People spoke more on correct trials (difference = 0.067, p<.0001), and both groups spoke more at the first success. But only the Same group kept speaking more afterward (post-success change = 0.074, p<.0001; Different was 0.007, not reliable). From content alone, correctness was classifiable at 71.7% accuracy, but Same vs Different on the same problem was not (49.2%, near chance).

The clearest sign of transfer was 'categorization' talk. In the Same group, explicitly naming the problem type rose from 0.3% to 2.1% after the first success — roughly a sevenfold increase (+1.8 points) — while the Different group stayed near zero. Meanwhile the number of distinct candidate moves fell at the first success in both groups (a sign of narrowed search), and fell further only in the Same group.

Use Cases

First, level ordering. If you are building a mechanic-teaching Sokoban-like or puzzle-platformer, placing consecutive levels that use the same trick lets players consolidate the knack before moving on — this design gains support. 'Introduce a new mechanic one at a time, repeat it for a while, then advance': that old maxim now has experimental backing on transfer.

Second, difficulty estimation. The paper assigns a difficulty ordering to matchstick transformation types (value-loose-chunk easy, tautological hard). If you are designing equation puzzles or 'break your assumption' one-move puzzles, that ordering can underlie your difficulty curve.

Third, giving mechanics a name. Since the sign of transfer was 'saying the problem type aloud,' letting players tag the trick they found — or gently prompting 'what kind of puzzle was that?' — may support transfer. It is a quiet way to gamify metacognition (reflecting on your own thinking).

Fourth, as playtest instrumentation: voice-activity detection, transcription, and semantic scoring can be repurposed into UX research that picks up, at scale, when an 'Aha' happens and whether transfer is occurring. Rising speech and narrowing candidate moves can serve as signals that a player is closing in. Fifth, hint timing: the finding that candidate moves narrow just before first success lets you conversely detect a player fanning out too many candidates as 'stuck,' and time hints to match.

Limitations

The authors are candid about limits. They use first success as a marker of re-framing but never directly measure a subjective Aha per trial. Think-aloud reflects only part of what is in the head, and automatic transcription and labeling add noise. Matchstick arithmetic is one tightly constrained insight task; other domains may show different verbal signatures. They suggest pairing speech with Aha ratings, testing whether early verbal features predict later transfer, and checking causally whether encouraging abstraction really increases reuse.

What I would add: this is a between-subjects design (two separate groups compared), so it does not track transfer within the same person, and individual differences may color the result. And while 'categorization' talk grew sevenfold, in absolute terms it is only 2.1% — a real effect, but a low base rate.

Further, speech rises alongside success; the paper does not show that verbalizing causes insight or transfer. This is a correlational observation, not causation. Using GPT-5.1 to label reasoning moves also warrants caution about measuring human reasoning with a machine's judgment (the authors say they hand-checked a subset). Finally, preregistered as it is, this is a single study and a pre-review preprint. I think it is wise to avoid generalizing to 'this is how people are,' and instead read it as 'under this task and these conditions, this was observed.'

Fukai's Reading

From here, my own interpretation. I would place this study between the classic cognitive-science view of 'insight as re-framing the problem' (Kaplan & Simon) and the craft wisdom that game design has accumulated — 'teach mechanics one at a time' — connecting them through the observation window of language. In the vocabulary of design criticism, it translates 'mastery through repetition of the same device' into speech, a signal you can measure from the outside. The point that 'being able to put the knack into words' marks transfer gives, I read, a quiet but real basis for the design choice of having players name the puzzle type — through the phrasing of tutorials and hints.

Closing

For those who want to go deeper: for the classic framing of insight as 'constraint relaxation and chunk decomposition,' read Knoblich et al. (1999); for the methodology of running think-aloud at scale — cited by this paper — read Wurgaft et al. (2025), 'Scaling up the think-aloud method.' For the neuroscience of insight, the Kounios & Beeman (2014) review is a good guide.

For my part, I drew a red line under the 'categorization talk grew sevenfold' paragraph in my printed copy. The moment a player mutters 'oh, this is the same as before' — that may be the very signal that the puzzle we built truly landed. I took a sip of strong coffee and thought about that.

References

Papers and related material referenced in this article:

・Leveraging Speech to Identify Signatures of Insight and Transfer in Problem Solving (Linas Nasvytis & Judith E. Fan, 2026, arXiv preprint arXiv:2605.12970)

・Related: Scaling up the think-aloud method (Wurgaft et al., 2025, arXiv preprint)

・Related: Constraint relaxation and chunk decomposition in insight problem solving (Knoblich, Ohlsson, Haider & Rhenius, 1999, Journal of Experimental Psychology: LMC)

・Related: The cognitive neuroscience of insight (Kounios & Beeman, 2014, Annual Review of Psychology)