Autistic traits foster effective curiosity-driven exploration

The study investigates how individual differences in autistic traits shape curiosity-driven exploration. Curiosity-driven exploration reflects intrinsic motivation to learn and involves selecting what, when, and how long to explore. Prior work shows that humans monitor learning opportunities and disengage when learning progress diminishes, but the relationship between such exploration drives and personality traits remains underexplored. Autistic traits, including insistence on sameness, represent a meaningful axis of variability that may alter tolerance to uncertainty and prediction errors. The authors test two competing hypotheses: (1) higher autistic traits may heighten motivation to reduce uncertainty via prioritizing learning progress—even when gains are small—leading to a stronger link between learning progress and exploration; (2) intolerance of uncertainty might drive avoidance of prediction errors, making exploratory decisions more guided by minimizing prediction errors rather than maximizing learning progress. The study examines when participants disengage from an ongoing activity and what they choose to explore next, relating these choices to trial-by-trial measures of prediction error, learning progress, and novelty, and to self- and other-reported autistic traits, with a primary focus on the insistence on sameness subscale.

Prior research indicates that curiosity and intrinsic motivation guide exploration toward activities with higher expected learning progress and novelty. Individual differences in exploration have been linked to personality traits (e.g., impulsivity, risk-taking), but mechanisms at the group level can obscure trait-related differences. Autistic traits have been associated with varied learning phenomena: reduced efficiency in volatile or probabilistic contexts and less robustness to noise in some studies, comparable performance to non-autistic groups in other tasks, and even enhanced statistical learning. Active learning and curiosity-driven exploration in relation to autistic traits have received little attention. A recent study showed that individuals with fewer autistic traits tolerated larger prediction errors before abandoning an environment. Given links between autistic traits, insistence on sameness, and intolerance of uncertainty, the current work examines whether higher autistic traits increase reliance on learning progress (to reduce uncertainty) or promote avoidance of prediction errors, and how these drives influence exploration and learning outcomes.

Ethics: Approved by Radboud University’s ethics committee (ECSW2016-0905-396). Digital informed consent obtained. Participants: 77 university students recruited; 7 excluded for 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 was not assessed. Procedure: Remote testing via Zoom. Task implemented in PsychoPy and hosted on Pavlovia; participants completed the task while on Zoom, then filled the Adult Social Behavior Questionnaire (ASBQ; self-report), task-involvement questions, and (if possible) arranged a parental other-report ASBQ. Task: In each of three settings (beach, sea, grassland), participants could freely choose one of four cartoon animals to engage with. On each trial, after selecting an animal, participants predicted where it would reappear from behind an occluder (rocks/waves/hedge). The animal’s true location was then revealed. Participants could switch animals at any time; previously explored animals remained available. Each animal could hide up to 35 times, or the setting advanced after 90 total plays. Total trials per participant ranged 150–271. No explicit instructions or external rewards directed learning. Hiding patterns: Each of the four animals in a setting followed a distinct generative pattern based on a Gaussian around a mean location with parameters manipulated to vary drift (gradual change), change-point probability (abrupt changes), and noise. Four pattern types per setting: (1) stable/low noise with small drift; (2) high noise (unlearnable); (3) high change-point probability with low noise; (4) high drift and high change-point probability with low noise. Pattern-to-animal mapping was counterbalanced; patterns were independent across animals. Behavioral variables: (a) Participants’ trialwise predictions (observable responses). (b) Leave-stay decisions (whether to disengage from the current animal). (c) Exploratory choices (which new animal was chosen after a leave). Questionnaires: ASBQ (44 items; six domains: reduced contact, reduced empathy, reduced interpersonal insight, violations of social conventions, insistence on sameness, sensory stimulation/motor stereotypies), self- and other-report versions. All participants completed self-report; 60% (42/70) provided parental other-reports. Analyses primarily relied on parental other-reports due to concerns about self-report reliability in higher autistic traits. Post-task questions assessed boredom timing (1st/2nd/3rd setting/never), favorites, and subjective experience. Data exclusions: participants reporting not understanding the task (N=7); for remaining participants, data collected after the reported onset of boredom were discarded (after 1st setting: 21 participants; after 2nd: 23; after 3rd: 26) before all analyses. Computational modeling: A Bayesian hierarchical delta-rule model with latent structure handled both gradual (drift) and abrupt (change-point) dynamics. Position estimate V was updated via a delta rule with an adaptive learning rate α* that probabilistically selected between a lower and a higher α each trial (indicator z_t ~ Bernoulli(ϕ)); drift V′ was learned via a separate delta rule with learning rate β. Responses B_t were modeled as Gaussian around V with precision η. Parameters (α_low, α_high, β, ϕ, η) were estimated per animal pattern and shared across individuals; trialwise latent states generated participant-specific estimates. Five variables derived per trial: prediction error (PE_t = |r_t − V_t|), learning progress (LP_t = PE_{t-1} − PE_t), novelty (N_t = −t, t = consecutive trials on the same animal), expected prediction error (delta-rule update from PE), and expected learning progress (difference between current PE and expected PE for next trial). Models were fit separately for each animal in each setting; only the selected animal’s model updated each trial. Fitting used JAGS (3 chains × 20,000 samples; 50% burn-in); convergence achieved (R̂<1.05). Model checks included simulations, parameter recovery, and comparisons (see S1 Text). Statistical analyses: Leave-stay decisions modeled with mixed-effects logistic regressions comparing four candidate models that included subsets of novelty, (expected) PE, and (expected) LP, all interacting with time (number of consecutive trials on an animal) and autistic traits (continuous), with random effects for participant, setting, pattern type, and animal. Best model selection based on other-reports, then replicated with self-reports. Exploratory decisions (choice among 3 remaining animals after leave) modeled with multinomial logistic regressions including expected learning progress, expected prediction error, and novelty; due to model limitations on cross-level interactions, traits were dichotomized by mean split and models run separately for low/high groups. Learning performance assessed via models of PE decline over trials across pattern types as a function of traits, comparing models with and without the three-way interaction (likelihood ratio tests).

• Leave-stay decisions (other-reports; best-fitting model included novelty, learning progress, and expected prediction error interacting with time and insistence on sameness):
  • Significant interaction LP × time × insistence on sameness: β = −0.15, SE = 0.05, p = 0.004.
  • Significant interaction expected PE × time × insistence on sameness: β = −0.10, SE = 0.05, p = 0.03.
  • Interpretation: Participants with lower insistence on sameness relied on learning progress early but shifted to expected prediction error later; participants with higher insistence on sameness showed greater persistence early and relied on learning progress later.
  • Novelty interaction was not significant (β = 0.06, SE = 0.05, p = 0.19).
  • Similar patterns for total autistic traits (LP × time × traits: β = 0.16, SE = 0.06, p = 0.004; expected PE × time × traits: β = −0.10, SE = 0.04, p = 0.03) and for reduced contact and reduced empathy subscales.
• Leave-stay decisions (self-reports):
  • Trend for expected PE × time × insistence on sameness (β = −0.07, SE = 0.04, p = 0.077).
  • Significant LP × insistence on sameness (β = −0.09, SE = 0.04, p = 0.01); LP × time × insistence not significant (β = −0.03, SE = 0.04, p = 0.47).
• Exploratory choices (multinomial logistic):
  • Other-reports (mean split; low N=17, high N=14): novelty predicted choices in both groups (low: β = 0.85, SE = 0.22, p < 0.001; high: β = 0.63, SE = 0.20, p = 0.002); expected PE and expected LP effects were in predicted directions but not significant (expected PE low: β = −0.15, p = 0.26; high: β = −0.24, p = 0.10; expected LP low: β = −0.07, p = 0.63; high: β = 0.19, p = 0.17).
  • Self-reports (mean split; low N=31, high N=26): novelty significant in both groups (low: β = 0.71, SE = 0.15, p < 0.001; high: β = 0.67, SE = 0.17, p < 0.001). High insistence group favored options with higher expected learning progress (β = 0.30, SE = 0.11, p = 0.006). Low insistence group avoided options with higher expected prediction error (β = −0.20, SE = 0.10, p = 0.047).
• Learning performance:
  • Time × pattern type × insistence on sameness (other-reports) significant: χ²(3) = 10.67, p = 0.014. Higher insistence on sameness associated with greater PE reduction over time across learnable patterns; no improvement for the high-noise (unlearnable) pattern.
  • Self-reports interaction not significant: χ²(3) = 6.32, p = 0.097. Total autistic traits (other-reports) also significant: χ²(3) = 8.53, p = 0.036; self-reports not significant: χ²(3) = 4.00, p = 0.26. Overall, individuals with higher insistence on sameness showed more persistence and leveraged learning progress later in exploration, yielding better learning on complex, probabilistic patterns. Individuals with lower insistence relied on learning progress early and on avoiding expected errors later, and preferred lower expected-error options when choosing what to explore.

The findings demonstrate that autistic traits systematically modulate curiosity-driven exploration. Regarding when to disengage, lower insistence on sameness was linked to early reliance on learning progress followed by a later shift to expected prediction error, indicative of switching from a growth-oriented drive to error avoidance over time. In contrast, higher insistence on sameness was associated with initial persistence (reduced reliance on either learning progress or expected errors) and later reliance on learning progress, reflecting a strategy of sustained engagement that capitalizes on learning gains as they accrue. In choosing what to explore next, all participants favored novelty, but trait-specific preferences emerged: individuals with lower insistence preferred options with lower expected prediction errors, whereas those with higher insistence prioritized options with higher expected learning progress. These patterns align with the hypothesis that insistence on sameness relates to sensitivity to learning progress rather than blanket avoidance of uncertainty. Importantly, these differing exploration drives translated into performance differences: higher insistence on sameness predicted greater improvements on learnable and more complex patterns (e.g., high drift), suggesting that persistence coupled with a learning-progress focus can be advantageous in complex environments. This stands in contrast to some previous findings portraying high autistic traits as impairments in volatile contexts; here, providing autonomy and intrinsically motivated exploration revealed strengths associated with higher insistence on sameness. Broader effects observed for total autistic traits and for subscales (reduced contact, reduced empathy) suggest shared underlying mechanisms or interrelations among traits that influence exploration strategy and learning outcomes.

This work shows that autistic traits, particularly insistence on sameness, shape how individuals explore and learn in curiosity-driven contexts. Lower insistence on sameness fosters early reliance on learning progress and later avoidance of expected errors, while higher insistence promotes persistence and later reliance on learning progress. When selecting new options, all participants favored novelty; additionally, lower insistence favored options with lower expected prediction error, whereas higher insistence favored options with higher expected learning progress. These trait-linked strategies were associated with better learning for individuals higher in insistence on sameness on complex, learnable patterns. The study highlights the importance of accounting for individual traits in models of curiosity-driven behavior and supports personalized learning approaches that leverage persistence and learning-progress sensitivity. Future research should develop analytic methods that preserve trait continuity in multinomial contexts, recruit larger and more diverse samples, and investigate neural/biological mechanisms linking autistic traits to exploration drives.

• Active, intrinsically motivated paradigm required substantial data curation: exclusion of participants who did not understand the task (N=7) and removal of data collected after self-reported boredom on subsequent settings, potentially affecting generalizability.
• Sample size constraints, especially for other-report mean-split multinomial analyses (small N per subgroup), may have limited power to detect effects in exploratory choices.
• Analytic limitations: multinomial logistic regression did not allow modeling trial-level predictors interacting with continuous between-subject traits, necessitating dichotomization (mean split) and loss of granularity.
• Online testing and variable total trial counts may introduce heterogeneity in engagement.
• Generalization is limited to university-aged participants; ASD diagnostic status was not assessed.
• Patterns were designed to induce variability rather than being the primary analytic focus; cross-animal parameter sharing across individuals in the model, while stabilizing estimates, may obscure some individual-specific parameter differences.