Developmental Change in the Triggers of Curiosity

• In Event: Fostering Children’s Early Curiosity, Engagement and Question-Asking Behavior

Abstract

Curiosity motivates people to search for information, and thus plays an important role in learning. Children are more likely to explore when they are surprised (Stahl & Feigenson, 2015) or uncertain (Schulz & Bonawitz, 2007), suggesting children’s curiosity is triggered by these cues. However, Liquin et al. (2020) argue that surprise and uncertainty are “heuristic” triggers of curiosity; if the goal of curiosity is to direct exploration towards situations that will likely lead to learning, then expected learning is the “optimal” trigger of curiosity. Is curiosity primarily sensitive to surprise and uncertainty, or does curiosity track expected learning? These cues are often correlated and have been confounded in prior work; we develop a paradigm to disentangle them.

Participants were 55 adults recruited through Prolific and 42 five- to nine-year-olds recruited from the local community (child data collection ongoing). The method was adapted from Dorfman et al. (2019). Participants chose one of two candy machines across 20 trials. Each machine, when chosen, produced “candy” with some probability. On 30% of trials, a hidden agent intervened, in which case the participant received candy with 100% probability (“donor” condition) or 0% probability (“thief” condition), regardless of the machine chosen. On each trial, participants rated how much they “wanted to know” about whether the hidden agent had caused the outcome (candy/no candy) on a 4-point scale.

We fit a Bayesian reinforcement learning model to the sequence of data observed by each participant. For each trial, the model estimated how much the outcome violated expectations (surprise), uncertainty about the chosen machine’s value (value uncertainty), expected learning about the chosen machine’s value (expected value learning), and expected learning about whether an intervention occurred (expected intervention learning). Replicating prior work (Liquin et al., 2020), only expected value learning (β=0.29) and expected intervention learning (β=0.32) explained significant variance in adults’ curiosity controlling for the other candidate triggers (ps<.001). In children, value uncertainty (β=0.12, p=.006) and expected value learning (β=0.15, p=.04) explained unique variance (Fig. 1), though expected value learning became a stronger predictor of children’s curiosity as age (in months) increased (β=0.02, p=.001). Comparing children’s and adults’ curiosity, value uncertainty was a stronger predictor of curiosity in children (interaction β=0.15, p=.002); expected intervention learning and surprise were stronger predictors of curiosity in adults (interactions β=-0.20, p=.008; β=-0.13, p=.03).

These results provide evidence for a developmental shift in the triggers of curiosity, potentially from heuristic cues to more optimal cues. Children’s curiosity most closely tracked uncertainty about the ultimate learning goal when controlling for other relevant considerations (though curiosity also tracked expected learning about the ultimate learning goal, increasingly across age). Adults’ curiosity tracked only expected learning, both about the ultimate learning goal and about trial-by-trial changes in the environment. Ongoing work using this paradigm has the potential to refine theories of curiosity and shed light on when and how curiosity motivates exploration and learning across development.

Authors

• Emily G Liquin, Princeton University
  Presenting Author

• Frederick Callaway, Princeton University
  Non-Presenting Author

• Tania Lombrozo, Princeton University
  Non-Presenting Author