Autistic traits foster effective curiosity-driven exploration

The study investigates how individual differences in autistic traits modulate curiosity-driven exploration. Curiosity-driven exploration refers to agents actively selecting what to explore based on intrinsic learning motives rather than extrinsic rewards. Prior work shows humans use learning opportunities—such as learning progress—to guide exploration. The core research questions are: (1) Do autistic traits alter the balance among exploration drives, specifically seeking novelty, minimizing prediction error, and maximizing learning progress? (2) How do autistic traits, particularly insistence on sameness, influence decisions about when to disengage from an activity and what to explore next? Two hypotheses are considered regarding intolerance of uncertainty associated with autistic traits: (a) individuals with higher autistic traits may prioritize reducing uncertainty via learning progress, even when small; predicting a stronger link between learning progress and exploration; (b) individuals with higher autistic traits may avoid uncertainty by minimizing prediction errors; predicting exploratory decisions guided by prediction-error avoidance rather than learning-progress pursuit.

Exploration levels are highly variable across individuals and relate to personality factors such as impulsivity and risk-taking, suggesting individual traits shape exploration mechanisms. Extensive research has examined sensory, cognitive, social, and communicative aspects of autistic traits, with mixed findings on learning: stronger autistic traits linked to less efficient learning about probabilistically aberrant events and reduced robustness to noise; comparable performance in visual search and decision-making tasks and in volatile environments; and reports of enhanced statistical learning abilities. Crucially, active learning and exploration have been underexplored in relation to autistic traits. A recent study showed that participants with fewer autistic traits tolerated larger prediction errors before leaving an environment, consistent with associations of autistic traits with insistence on sameness and intolerance to errors. How these traits influence the balance among drives—novelty seeking, prediction-error minimization, and learning-progress maximization—remains unknown, motivating the present investigation.

Ethics: Approved by Radboud University’s ethics committee (ECSW2016-0905-396); digital written consent obtained. Participants: 77 university students recruited; 7 excluded due to not understanding the task; final N=70 (14 men, 51 women, 5 nonbinary; age 17–35, M=22.2, SD=4.2). ASD diagnosis status not assessed. Procedure: Remote testing via Zoom; task built in PsychoPy and hosted on Pavlovia. Post-task, participants completed the Adult Social Behavior Questionnaire (ASBQ) self-report and, if possible, the other-report (parental figure). Additional task-experience questions were collected. Task: Across three settings (beach, sea, grassland), participants freely chose among four cartoon animals and could switch at any time. On each trial, selecting an animal triggered its hiding behind a row of locations; participants clicked to predict the reappearance location, then received feedback on the true location. The game advanced after any animal had hidden 35 times or the participant had played 90 trials per setting. Total trials per participant ranged 150–271. No external rewards or cues were provided. Hiding patterns: Each animal followed a Gaussian-based hiding pattern defined by mean location, noise (SD), drift (slow movement over trials), and change-point probability (abrupt changes). Four pattern types per setting: (i) low noise, small drift, low change-point probability (learnable); (ii) high noise, no drift, no change points (unlearnable); (iii) high change-point probability, low noise, no drift (learnable); (iv) high drift, high change-point probability, low noise (learnable). The mapping of pattern to animal was counterbalanced across participants. Measures: (1) Trial-wise participant predictions (click location) used to fit the model. (2) Leave-stay decisions: whether participants disengaged from the current animal. (3) Exploratory decisions: which of the remaining animals was chosen after a leave. Questionnaire: ASBQ (44 items; six domains: reduced contact, reduced empathy, reduced interpersonal insight, violations of social conventions, insistence on sameness, sensory stimulation and motor stereotypies) using 3-point Likert scale; both total and subscale scores computed. 100% provided self-reports; 60% (42/70) provided parental other-reports. Analyses prioritized other-reports due to reliability concerns, but results for both report types are presented. Data following self-reported boredom were excluded (after setting 1: 21 participants; setting 2: 23; setting 3: 26) to focus on intrinsically motivated exploration. Computational modeling: A Bayesian hierarchical delta-rule model captured learning of both gradual (drift) and abrupt changes via adaptive learning rates. The model updated belief about position by integrating prediction error with a trial-wise learning rate that could switch between low and high values (regulated by a Bernoulli indicator), and separately tracked drift via a delta-rule on velocity. Response noise was modeled as Gaussian variance. Parameters (two alphas for fast/slow updates, phi for switch probability, beta for drift learning, eta for response noise) were estimated per animal across individuals. Five key variables were derived: unsigned prediction error (PE), learning progress (LP = decrease in PE across consecutive trials), novelty (inverse of exposure/time with an option), expected prediction error (delta-rule forecast of next-trial PE), and expected learning progress (derived from the difference between expected and actual PE). Models were fit in JAGS with three chains, 20,000 samples each (first half burn-in); convergence achieved (R<1.05). Trial-by-trial variables were computed only for the currently engaged animal; others remained unchanged. Statistical analyses: Leave-stay behavior was modeled via mixed-effects logistic regression comparing models including subsets of novelty, (expected) prediction error, and (expected) learning progress, each interacting with time (consecutive trials with the same animal) and autistic traits (treated continuously). Random effects: participant, setting, pattern type, animal type. Model selection based on AIC/BIC/log-likelihood. Exploratory choices (multinomial outcomes among three options post-leave) were analyzed with multinomial logistic regression including expected learning progress, expected prediction error, and novelty. Because multinomial models did not permit interactions with continuous between-subject traits, mean splits on autistic traits were used to form high/low groups. Performance analyses tested prediction error reduction over time by pattern type with autistic traits as a continuous moderator; models with and without the three-way interaction were compared via likelihood ratio tests.

Leave-stay decisions (other-reports, insistence on sameness): The best-fitting model included novelty, learning progress, and expected prediction error interacting with time and insistence on sameness (AIC=3657, BIC=3772, Log-Likelihood=-1812). Significant interactions: learning progress × time × insistence on sameness (β = -0.15, SE = 0.05, p = 0.004) and expected prediction error × time × insistence on sameness (β = -0.10, SE = 0.05, p = 0.03). Novelty interaction was not significant (β = 0.06, SE = 0.05, p = 0.19). Pattern: Low insistence on sameness participants relied on learning progress early and shifted to expected prediction error later; high insistence on sameness participants were more persistent early and relied on learning progress later. Full-scale autistic traits (other-reports): Similar significant interactions for learning progress × time × traits (β = 0.16, SE = 0.06, p = 0.004) and expected prediction error × time × traits (β = -0.10, SE = 0.04, p = 0.03). Subscales reduced contact and reduced empathy showed similar patterns. Self-reports: Trend-level expected prediction error × time × insistence on sameness (β = -0.07, SE = 0.04, p = 0.077). Significant learning progress × insistence on sameness (β = -0.09, SE = 0.04, p = 0.01), but LP × IoS × time not significant (β = -0.03, SE = 0.04, p = 0.47). Exploratory decisions: Novelty robustly increased choice probability across high/low groups and report types (e.g., self-report low group β = 0.71, SE = 0.15, p < 0.001; high group β = 0.67, SE = 0.17, p < 0.001). Self-reports: High insistence on sameness group selected options with higher expected learning progress (β = 0.30, SE = 0.11, p = 0.006); Low insistence on sameness group preferred options with lower expected prediction error (β = -0.20, SE = 0.10, p = 0.047). Other-reports showed effects in the same directions but did not reach significance (expected learning progress high group β = 0.19, SE = 0.14, p = 0.17; expected prediction error low group β = -0.15, SE = 0.14, p = 0.26). Learning performance: The three-way interaction of pattern type × time × other-reported insistence on sameness was significant (χ²(3) = 10.67, p = 0.014). Higher insistence on sameness related to greater prediction error reduction over time across learnable patterns (stable, high CPP, high drift), but not in the unlearnable high-noise pattern. Self-reports did not reach significance (χ²(3) = 6.32, p = 0.097). Full-scale autistic traits (other-reports) showed similar performance benefits (χ²(3) = 8.53, p = 0.036).

Findings demonstrate that autistic traits, particularly insistence on sameness, shape curiosity-driven exploration. Individuals lower in insistence on sameness leave activities early when learning progress wanes and later switch to avoiding high expected errors, suggesting a utility function that prioritizes error minimization. Individuals higher in insistence on sameness show greater persistence early and later rely on learning progress to disengage, and in choice favor options promising higher learning progress—suggesting a utility function that prioritizes learning. These drives translate into improved performance for those higher in insistence on sameness on learnable, probabilistic, and complex (high drift) sequences, highlighting persistence as advantageous in tasks requiring sustained engagement. Results extend beyond insistence on sameness to broader autistic traits (e.g., reduced contact, reduced empathy), suggesting shared underlying mechanisms or causal relations among traits. The study underscores the importance of accounting for individual differences to avoid confounds in group-level analyses of exploration and provides evidence that enabling self-paced, self-selected exploration can reveal strengths associated with autistic traits.

Autistic traits modulate both when learners disengage and what they choose to explore. Lower insistence on sameness is associated with early reliance on learning progress and later avoidance of high expected errors, while higher insistence on sameness is associated with persistence and later reliance on learning progress, as well as preferential selection of options with higher expected learning progress. These patterns yield superior performance for those higher in insistence on sameness on learnable patterns. The work highlights the value of personalized, curiosity-aligned learning environments and reframes certain autistic traits as strengths in exploration contexts. Future research should elucidate neural mechanisms linking autistic traits to exploration drives, examine causal relationships among traits, and develop analytic tools that accommodate continuous trait measures in multinomial choice models.

Active, intrinsically motivated tasks require data curation: several participants were excluded due to misunderstanding instructions, and trials after self-reported boredom were removed, potentially affecting generalizability. Sample sizes for some subgroup analyses (especially other-report mean-splits) were modest, limiting power. Multinomial logistic models necessitated mean splits on traits, reducing granularity; methods that integrate continuous individual differences into multinomial choice models are needed. Findings pertain to university students and may not generalize to broader populations or diagnosed ASD groups.