The research team recruited 127 adolescents aged 10 to 20 years to complete a social prediction game. In each trial, participants watched two players on opposite sides of a machine, each holding an endowment of five coins. The players could choose to keep or share their coins, and the machine displayed all possible outcomes based on both choices. Two outcome values, the temptation to defect and the sucker’s payoff, varied systematically across 16 games drawn from four types of economic games. Participants predicted the choice of one player, who remained the same across 16 trials, and then rated their confidence. Each player followed a pre-programmed strategy that produced a unique response pattern across the game space, and each player cooperated on exactly half of the trials, so accuracy above 50 percent would indicate genuine learning and generalization.

The behavioral results confirmed that participants learned more than surface patterns. Average accuracy across the task was 58.5 percent, significantly above the 50 percent chance level, with a standard deviation of 22.3 percent. More strikingly, participants were far more accurate when predicting players whose behavior followed human motives than when predicting players following artificial, non-human motives. Accuracy reached 74.5 percent for human-motive conditions but only 42.4 percent for artificial-motive conditions, a mean difference of 32.1 percentage points. Among the human-motive players, accuracy was 73.7 percent for the greedy player and 75.3 percent for the risk-averse player, compared with 41.6 percent and 43.4 percent for the inverse versions of those motives. Participants also became more accurate over the course of the task, with an odds ratio of 1.17 per unit of time, and older participants were significantly more accurate than younger ones, with an odds ratio of 1.11. The interaction between age and time was not significant, suggesting that the learning trajectory itself was similar across ages even though older adolescents started from a higher level of performance.

The central question was what kind of mental structure participants used to make their predictions. The researchers fitted a series of computational models, each representing a different generalization strategy. A no-structure model updated beliefs by directly associating specific actions with an individual, while a game type-based model organized knowledge by the type of economic game. The motivation-based models instead identified latent variables, such as motives, that consistently generate behavior across contexts. The full motivation-based model, which included all four motives tested, outperformed both the no-structure model and the game type-based model, winning the pairwise comparison for 122 and 115 of the participants respectively. The best fit of all came from a subset variant that allowed each participant to consider only a subject-specific subset of the four motives, with a mean Bayesian information criterion of 23.9 compared with 35.3 for the full model. A random-effects Bayesian model selection analysis strongly favored the subset model, with a protected exceedance probability of approximately 1.00. The result held when the analysis was repeated using the sum of squared errors instead of the BIC, confirming that the advantage was not an artifact of the complexity penalty.

Examining which motives participants included in their best-fitting models revealed a distinctive developmental pattern. The most commonly inferred motive was expected value maximization, included for 64 percent of participants, followed by risk aversion at 49 percent, greed at 45 percent and cooperativeness at 24 percent. This pattern differs markedly from what the earlier adult study by van Baar and colleagues had found, where greed was the most prevalent inferred motive, considered by 87 percent of adults, while only 21 percent considered risk aversion. The authors suggest this difference may reflect an age-related shift in the baseline from which people evaluate social decisions. In the game, the expected value strategy is to cooperate unless the temptation to defect exceeds the sucker’s payoff by five coins, whereas the greedy strategy tolerates defection until the temptation exceeds ten coins. Adolescents may therefore compare the temptation to defect against the outcome of mutual mistrust, while adults anchor on the outcome of mutual cooperation.

Age shaped not only which motives were inferred but how many. The likelihood of considering more than one motive increased with age, meaning that the probability of relying on only a single motive declined with a coefficient of -0.41. Age was not significantly associated with considering more than two or three motives. When the researchers explored which specific motive drove this developmental change, they found that the probability of inferring risk aversion increased with age, particularly from early to mid-adolescence, with a linear estimate of 1.63 and an odds ratio of 5.10. The probabilities of inferring greed, expected value or cooperativeness were not associated with age. The authors interpret this as partial confirmation of their main hypothesis: the ability to infer the motives of others improved with age, and this improvement was driven particularly by better inference of risk aversion from early to mid-adolescence.

The study also tested whether adolescents project their own motives onto others, a bias known as self-anchoring. Among the 127 participants, 20 followed the risk-averse motive in their own choices, 16 the expected value motive, 16 the greed motive and 12 the cooperativeness motive, while 63 did not match any specific motive pattern. Contrary to the pre-registered expectations, self-anchoring did not significantly improve prediction accuracy for risk-averse or greedy players, and the interaction between self-anchoring and age was also not significant. The authors note that null findings should be interpreted with caution, since they may depend on task demands or perceived similarity between participants and the observed agents, but they suggest that adolescence may be a period in which people increasingly differentiate their own perspective from that of others, which could be adaptive for learning that others think and value things differently.

Individual differences told a different story than expected. The researchers found no significant relationships between mental model complexity and environmental harshness, social anxiety or cognitive empathy, and no significant relationship between social anxiety and being risk-averse. The authors suggest that the expected relationships may emerge mainly in self-relevant contexts that engage affective processes, since the social prediction task required participants to observe others from a third-party perspective without being personally affected by the decisions. They also emphasize that third-party observation is itself an ecologically valid mode of social learning, noting that much of what adolescents learn about social norms, reputations and other people’s character comes from observing interactions they are not directly part of, such as peer conflicts or group dynamics.

The findings extend prior developmental research that has mostly examined how adolescents learn what decisions another person typically makes in repeated, identical settings. By showing that adolescents build higher-order representations of why others act as they do, the study positions motive inference as a higher-level component of adolescent social learning that goes beyond context-independent action-outcome associations. The authors note that the within-subjects manipulation of human and artificial motives, the randomization of trial and block order, and the computational modeling approach all strengthened the design, although data were collected online during the later part of the COVID-19 pandemic, which could have affected social experience. Because the design was cross-sectional, the study cannot establish within-person developmental change, and the authors call for longitudinal studies with large samples to test whether adolescents become increasingly able to infer multiple motives over time and how such changes relate to real-world social experience.

For now, the study offers a computational window into a skill that develops quietly through everyday observation. Understanding how these latent motive representations form, the authors argue, will be essential for refining theoretical models of adolescent social development and may eventually inform interventions aimed at fostering social competence during a sensitive period for social learning.

Subject of Research: Developmental computational study of how adolescents infer latent motives from observed social decisions

Article Title: Adolescents infer latent motives to build generalizable models of others’ behavior

Article References: Vigil-Escalera, J., van Baar, J., Molleman, L., van Duijvenvoorde, A. C. K., & Ma, I. (2026). Adolescents infer latent motives to build generalizable models of others’ behavior. iScience, 29(11), Article 118039. https://doi.org/10.1016/j.isci.2026.118039

Image Credits: AI Generated

DOI: 10.1016/j.isci.2026.118039

Keywords: adolescence, social learning, motive inference, computational modeling, reinforcement learning, risk aversion, greed, cooperativeness, economic games, social prediction, developmental psychology, decision making