Experimental study · 2026
Pupil dilation as an alternative signal for adaptive memory modelling
An eye-tracking study comparing pupil dilation with response time as inputs to an adaptive memory model, with individual analysis of learning and scheduling.
Study design documented in the individual report; each participant completed both conditions.
Individual contribution
My role
I contributed to implementing the experiment software and running experimental sessions. The five-person team shared the study design, stimuli and data collection. I independently carried out the analyses, interpreted the results and wrote the individual report linked here.
Contribution statement in the reportResearch question
Can task-evoked pupil dilation replace response time as an input to SlimStampen, an adaptive memory model? The study examines learning outcomes, whether each signal tracks estimated memory strength, and how it affects the scheduling of practice. Accuracy remains an input in both conditions.
Study design
Eleven participants completed both conditions in a counterbalanced experiment at the University of Groningen: 22 sessions, with 90 learning trials per session and a four-alternative knowledge test. The experiment used OpenSesame, PyGaze and an EyeLink Portable Duo. The pupil condition included a three-second response-free retrieval window. Analyses used exact permutation tests and replayed each session through the model to examine memory estimates and schedules.
Findings
Knowledge-test accuracy on items presented during learning averaged 93.5% in the pupil condition and 85.3% in the response-time condition. The 8.2 percentage-point difference was not statistically significant in the full sample (exact permutation p = .086). In the nine participants tested after matched retention intervals, the difference was 12.0 points (p = .031); this smaller-subset result should be read alongside the full-sample finding. Pupil responses did not track the model’s estimated memory strength, and the two models produced similar practice schedules.
Interpretation and limitations
The experiment demonstrates that pupil measurements can drive the scheduling pipeline, but does not establish a learning benefit attributable to the pupil signal. The additional three seconds of retrieval practice is a confound. Interpretation is also limited by the small sample, high accuracy, unusable pupil measurements and model parameters taken from the literature rather than fitted to these data. A feedback software fault extended study time in both conditions. A follow-up would need equal retrieval windows and a more demanding task to isolate the signal’s contribution.
Related research
AI coding assistance & maintainability