Missing Vowels, Lasting Words: Tracking the Effort of Generation

A word with its vowels removed asks the reader to supply something. The remaining letters and a related cue may make the answer recoverable, but recovering it is an activity, not simply another way of looking at a completed word. This difference has long interested memory researchers. If generated words are remembered better, is the benefit explained by the extra effort involved? Answering that question requires more than comparing final test scores. Otherwise, the explanation risks becoming circular: an activity was effortful because it produced better memory, and memory improved because the activity was effortful.

Ania M. Grudzien and Nash Unsworth approached the problem by recording pupil responses while people learned word pairs. Their paper appeared online in September 2025 and in a 2026 issue of Memory & Cognition. Across four experiments, it connected a physiological measurement made during learning with a later measure of cued recall. The most revealing findings were not cases in which everything rose together. Some comparisons separated the pupil response from the memory benefit, making a simple equation between working harder and remembering more difficult to sustain. [1]

An answer assembled during its first presentation

In the first experiment, forty-one University of Oregon participants studied pairs consisting of a cue and a target. Targets in the reading condition appeared complete. In the generation condition, vowels were replaced with blanks, so participants had to work out the missing word. Generation occurred during encoding, with the cue and remaining letters available. It was not a manipulation in which participants first studied an intact answer and later practised retrieving it without access to the material.

Two practice rounds preceded three experimental rounds. Each experimental list contained twenty-four pairs, divided equally between reading and generation. A fixation cross appeared for two seconds, followed by a word pair for five seconds. After the list, cues returned individually and participants had five seconds to type each target. Both learning conditions were therefore experienced by the same person. The task could compare what that person later remembered from completed and incomplete targets, while recording what happened during the original encounter. [1]

A response relative to a baseline

The eye tracker recorded pupil diameter during learning. The analysis examined changes relative to the baseline before each pair, rather than treating naturally large pupils as a sign of superior memory. Measurements were grouped into two-hundred-millisecond intervals across the five-second encoding period. Missing observations associated with blinks, off-screen fixations or equipment problems were omitted from averaging rather than filled in through interpolation.

Generated targets produced better recall in this first experiment, alongside larger task-evoked pupil responses. The pattern was consistent with greater attentional effort during generation. That interpretation nevertheless requires a distinction between an indicator and a direct reading of a mental state. Pupil diameter responds to light and emotional influences as well as cognitive activity. Additional analyses excluded the early light-reflex portion of the response, but no pupil measurement can by itself identify the exact thought a person is having. Nor does a larger response automatically explain why a later answer is correct. [1]

When extra effort had no demonstrated memory payoff

The second experiment moved the manipulation between people. Eighty-two participants were divided into two groups of forty-one. One group read complete targets throughout, while the other generated targets throughout. Individuals no longer moved between the two learning conditions within their own lists. The generation group again showed a larger pupil response, suggesting that the task demanded more attentional effort.

This time, recall accuracy did not significantly differ between the groups. Both means rounded to the same value. That is not a proof that generation can never help in a between-person comparison, or a universal demonstration of equivalence. It is evidence that the physiological difference observed here did not bring the predicted behavioral advantage. A general account must accommodate that separation, rather than describing the pupil result as if it already contained a successful memory outcome.

The authors considered whether contrasting activities within one task changes how attention is distributed. Generated items might gain an advantage relative to read items partly because the learner allocates resources differently across them. Someone who only generates does not experience that same local contrast. This remains an explanatory possibility, not a mechanism conclusively isolated from every difference between the separate experiments. The important observation is narrower and firmer: the study could detect a larger pupil response without detecting better recall in that comparison. [1]

The easier condition challenged the prediction

Fifty participants took part in a third experiment with reading, easier generation and harder generation conditions. Difficulty depended on associative relationships. Easier targets were strong associates of their cues; harder targets were less strongly associated. New word pairs supplied the easy condition. Thus, the manipulation did not simply take the same targets and remove different numbers of letters. The relationship between cue and answer changed as well.

The prediction was that difficult generation would recruit more effort and yield the strongest memory. Instead, easy generation produced the highest recall, followed by difficult generation and then reading. The easy condition also tended to produce a somewhat larger and more sustained pupil response. Increasing the nominal difficulty of generation did not arrange the results in the predicted order.

One complication is that a strongly associated target can be guessed at test. A participant might arrive at the correct answer from the cue without having successfully generated it during encoding. In that case, a correct response does not uniquely establish memory for the earlier learning episode. The authors also considered whether finding a target could produce an orienting response that contributes to pupil dilation. These possibilities matter because they prevent the easy condition from becoming a simple demonstration that lower difficulty causes greater effort or better learning. Several comparisons of the pupil curves were exploratory or reviewer-suggested, and different ways of defining peaks did not produce identical significance outcomes. [1]

Recording what the learner actually produced

The fourth experiment added a way to inspect the generated content. Forty-six participants said both cue and target aloud in the reading and generation conditions, while a research assistant recorded their responses. Speech was therefore not an extra action restricted to generated items. It let the researchers establish which word the participant had assembled during encoding.

Participants generated the experimenter-intended target on about seventy-seven percent of generation trials. That figure describes successful target production during study, not later recall accuracy. Subsequent answers were scored in two ways. One scheme required the intended target. The other accepted the word the participant had actually spoken earlier, even if it differed from the intended target. Both scoring methods retained a generation advantage.

This distinction makes the test more informative. A response can fail to match the official target while still reproduce the learner's actual earlier answer. Conversely, a target supplied at test may not prove that the same target was successfully produced before. Recording the encoding response helps separate these possibilities. It does not make every remaining answer unambiguous, but it improves the connection between what happened during learning and what the later score represents. [1]

What the combined evidence supports

An exploratory analysis across the within-person experiments statistically accounted for differences in the pupil response. The estimated behavioral generation effect became smaller, but remained significant in the pooled data. This supports a partial relationship between effort-related processes and the memory advantage. It does not establish that enlarging the pupil causes better memory, or that attentional effort supplies the whole explanation. The analysis was not a separate intervention that manipulated effort while holding every other process constant.

The paper also addressed the dependence between neighboring pupil measurements: a pupil cannot change size instantaneously, so adjacent time intervals are not independent samples of behavior. Additional models incorporated that temporal structure. Their results offered support for differences in response trajectories, but did not make every main effect significant in every experiment. Treating all analyses as identical confirmations would erase some of the statistical detail that makes the measurement interpretable.

The studies were not preregistered, used university participants, and allowed only five seconds for typed recall, which could leave incomplete answers and scoring problems. Their classroom implications remain proposals rather than demonstrated long-term educational outcomes. A useful conclusion is therefore not to make all learning harder. It is to examine what learners actually generate, how their attention is measured, and what a correct test response can establish. Missing vowels create a problem to solve; the pupil records part of the response to that problem; recall assesses a later outcome. Keeping those stages distinct allows effort to be investigated without turning effort into a promise. [1]

Sources

[1] Grudzien, A. M., and Unsworth, N. (2026; first published online 2025). Investigating a mental effort explanation of the generation effect using pupillometry. Memory & Cognition, 54, 980–996. https://link.springer.com/article/10.3758/s13421-025-01791-0

The source study is licensed under Creative Commons Attribution 4.0: https://creativecommons.org/licenses/by/4.0/ This article is an original explanatory account and assessment of the research, not an official translation or reproduction of the paper.