Prediction was the clearest dividing line
Partners whose predictions about the consequences of their own actions were more closely aligned tended to coordinate better, even after individual skill was taken into account. The finding is an association within this task: it does not establish that aligning those predictions would itself improve coordination.
After quality and score exclusions, the final analysis included 70 participants contributing 64 valid dyads. The exclusions removed 9 of 91 individual-action tasks and 15 of 79 joint-action tasks. The criteria included an eye-tracking quality score below 65% or a total score below 55%.
Participants completed both individual and joint tasks. Mean event success was 90.82% in the individual task and 85.21% in the joint task. Rather than compare raw joint scores alone, the researchers used a Bayesian generalized linear model, or GLM, based on the mean of partners' individual-task scores to set an expected level, then labeled pairs above or below that expectation.
Eye movements carried much of the signal
To look for useful predictors, the researchers used XGBoost, a machine-learning classifier, to combine individual-task gaze, motor, heart-rate, movement-energy and static predictors. The pair-level evaluation used stratified threefold cross-validation, with 67% allocated to training and validation and 33% held out for testing.
Among models built from separate feature categories, the full model had the highest accuracy and MCC, while eye movement ranked first as a single category. Seven of the 15 top predictors measured eye movement, four captured contact, two measured movement intensity, one came from the pre-contact swing and one from heart rate. No static-characteristic predictor appeared in that top group.
In a stripped-back test, a parsimonious, or deliberately smaller, model using only the two top predictors classified pairs with 75% accuracy. The reported 95% confidence interval was 73.66% to 76.34%, and the two-predictor model slightly outperformed the full model by 4.69%.
The stronger pairs watched and moved differently
The researchers then compared above- and below-expected pairs with hierarchical Bayesian t-tests. These tests examined anticipatory gaze, movement dynamics and static characteristics. Corrective-saccade counts used a Poisson likelihood, while continuous pre-contact predictors used a skewed Student-t distribution.
Above-expected pairs showed credible differences in eye movement, the swing before contact, contact behavior and positional alignment. They did not differ credibly in static characteristics, heart rate, movement intensity or self-assessment. Both hitters and receivers made fewer corrective saccades, or brief gaze adjustments, in both phases. The median standardized differences were -0.58 for hitters and -0.45 for receivers in Phase 1, and -0.47 and -0.48 in Phase 2, respectively.
A separate generative model, designed to represent how observations could arise, examined visual-angle error in Phase 1 using 3,330 observations, including 1,749 from above-expected and 1,581 from below-expected pairs. The model combined prior expectations with sensory evidence inferred within the model.
Both sources entered the model, but their weights differed sharply. The median weight placed on sensory evidence was 0.37 for above-expected pairs, with a reported 95% highest-density interval of 0.27 to 0.47, compared with 0.81 for below-expected pairs, with an interval of 0.75 to 0.87. In practical terms, the lower weight in the above-expected group means those predictions relied more on prior expectations than on sensory evidence.
The evidence has a narrow scope
The study supports an association within this task, not a cause-and-effect claim. It does not establish that aligned self-predictions cause better coordination, nor that the same pattern would hold in other settings.
The task was constrained and brief, so the findings may not carry over to long-term cooperation, communication-rich interaction, physically interdependent work or larger groups. Some participants also contributed to more than one dyad, creating dependence among observations.
The study reports ethics approval and written informed consent. It was not preregistered, and part of the individual-action dataset had been reported previously; the sample-size justification was referred to in supplementary information.
Deidentified participant data and analysis code were stated to be publicly available under a CC BY 4.0 license, although the supplied material did not include repository locations. Supplementary material is listed as available online. Funding was attributed to CASCB through the DFG Excellence Strategy; funders were reported to have had no role in the study, and the authors declared no competing interests.
Paper data and sources
Original title: Successful coordination emerges from aligned self-predictions in a dynamic motor interaction.
Authors: Prasetia Utama Putra, Fumihiro Kano
Journal/Repository: Communications psychology
Status: Peer-reviewed
First online: 2026-08-20
DOI: 10.1038/s44271-026-00523-7
Original paper · Full text