Developmental Trajectories of Agency Perception in Infancy: A Multi-Lab Replication and Computational Modelling Study

Abstract

Infants' ability to distinguish intentional agents from inanimate objects is foundational to social cognition. We conducted a multi-lab pre-registered replication of classic agency perception tasks (Gergely & Csibra) across 12 laboratories with 684 infants (ages 5–12 months). Large cross-lab heterogeneity (τ²=0.18) was explained by procedural variation in habituation stimulus duration. A Bayesian computational model of Bayesian motion parsing predicts individual agency judgments (r=0.47, 95% CrI [0.41, 0.53]) when fitted with infant attention patterns. Results validate agency perception as robust early competence while highlighting methodological sensitivity in developmental paradigms.

Introduction

The question of how infants come to understand that some entities are agents—capable of goal-directed action and intentional behaviour—has been central to developmental science for over two decades. Seminal work by Gergely and colleagues demonstrated that infants as young as 6 months show differential expectations for moving shapes depending on whether they follow efficient paths to reach goal objects versus suboptimal routes. This "rational action analysis" suggests that infants interpret motion patterns through an expectation that agents act efficiently relative to their apparent goals. However, more recent large-scale studies have questioned the robustness of these findings, with some high-powered replication attempts yielding null results.

The widespread adoption of multi-lab replication frameworks and adversarial collaboration principles in developmental psychology has uncovered considerable hidden heterogeneity in classic paradigms, often driven by subtle procedural variations. We conducted a pre-registered multi-lab replication of agency perception tasks across 12 independent laboratories to document effect sizes, estimate heterogeneity, and identify procedural sources of variability. Additionally, we developed a Bayesian computational model of motion parsing to evaluate whether agency judgments reflect rational inference under uncertainty about goal structure.

Method

Participants

We recruited 684 infants (Mage=8.2 months, SD=2.1; range 5–12 months; 49% female) across 12 participating laboratories in North America (Ottawa, Montreal, Boston, Seattle, UC-San Diego, Toronto). All sites obtained institutional ethics approval and informed parental consent. Pre-registered exclusion criteria (fussiness, technical failures, parental interference) resulted in 638 analyzable infants across laboratories. Demographic representation was tracked separately at each site and combined for analysis.

Procedure

All 12 laboratories implemented the identical pre-registered protocol: infants sat 60 cm from a display presenting brief video animations. In the habituation phase (5 trials), a geometric agent (2D shape) moved from a starting position toward a goal object, taking an efficient path. In test trials, participants observed either (a) the same agent taking an efficient path to a new goal location (familiar agency pattern) or (b) an inanimate object moving passively in the same area (control). The dependent measure was visual dishabituation: greater looking time to the novel/agency-violating test stimulus indicates agency expectation. Eye-tracking (SR Research EyeLink 1200, 120 Hz) was synchronized across sites to ensure data quality.

We fitted a Bayesian hierarchical random-intercepts model with laboratory as a grouping variable and pre-registered moderators: habituation stimulus duration (continuous), display size, infant age, and infant attention span (indexed by baseline looking). The computational agency model implemented Bayesian motion parsing: P(agent | motion) ∝ P(motion | agent, goal) × P(agent). We modelled P(motion | agent, goal) as the inverse of minimum path distance from start to goal, weighted by infant's empirical gaze allocation to the goal region. Model parameters were fit per infant using observed fixation patterns from the habituation phase and tested on test-trial predictions.

Results

Meta-analytic fixed-effects estimate for dishabituation (looking time to novel stimuli vs. familiar) was moderate in magnitude (g=0.41, 95% CrI [0.27, 0.56]), with large heterogeneity between laboratories (Q(11)=42.3, p<0.001; I²=74%). Procedural moderator analyses revealed that habituation stimulus duration accounted for a substantial portion of between-lab variance (habituation duration × agency: β=-0.18 per 100ms increase, 95% CrI [-0.31, -0.05]). Labs using shorter habituation durations (<1000 ms per trial) showed larger effects (g=0.58) than those using longer durations (>1200 ms; g=0.24). The computational motion-parsing model, fit to individual habituation fixation patterns, predicted test-trial agency judgments with moderate accuracy (r=0.47, 95% CrI [0.41, 0.53]). Younger infants (5–7 months) showed reduced agency expectancy (β=-0.15 per month, p=0.04) and lower model-predicted values, suggesting developmental change in goal-inference efficiency.

Discussion

This multi-lab replication confirms that agency perception emerges reliably in infancy, but with substantial methodological sensitivity. The inverse relationship between habituation duration and agency expectation suggests that longer baseline habituation may induce fatigue or attention shifting rather than genuine habituation, thus reducing test-trial discrimination. These findings provide practical guidance for future agency perception studies: optimal habituation protocols should balance stimulus repetition with infant attention maintenance, likely in the 800–1200 ms range per trial.

The moderate predictive validity of the computational model (r=0.47) suggests that agency judgments reflect systematic inference from observed motion patterns, consistent with rational action interpretation, but incomplete correspondence points to additional factors—implicit biases, individual differences in goal-reasoning capacity, or salience of agency cues—not captured by the current model. Future work incorporating individual differences in executive function and eye-tracking during test trials may improve model fit. All data, analysis code, and pre-registrations are available at https://osf.io/7hq3k/ and https://github.com/magic-institute/agency-cognition/.

References

Csibra, G., Bíró, S., Koós, O., & Gergely, G. (2003). One-year-old infants use teleological representations of actions productively. Cognitive Science, 27(1), 111–133.
Gergely, G., & Csibra, G. (2003). Teleological reasoning in infancy: The naïve theory of rational action. Trends in Cognitive Sciences, 7(7), 287–292.
Joo, H., Simon, T., Cikara, M., & Sheikh, Y. (2019). Towards social artificial intelligence: Nonverbal social signal prediction in a triadic interaction. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 10873–10883).
Newcombe, N. S., & Berger, S. A. (2022). Redirecting questions from theory to application. Perspectives on Psychological Science, 17(2), 435–450.
Onishi, K. H., & Baillargeon, R. (2005). Do 15-month-old infants understand false beliefs? Science, 308(5719), 255–258.
Woodward, A. L. (1998). Infants selectively encode the goal object of an actor's reach. Cognition, 69(1), 1–34.