Humans often make decisions based on information that is irrelevant to actual outcomes, a tendency widely documented in behavioral economics and cognitive psychology as a hallmark of biased or suboptimal learning. While prior research has shown that individuals with autism may be less susceptible to such biases in structured, static decision-making contexts, it remains unclear whether this resistance generalizes to more complex, dynamic learning environments. In this pre-registered adversarial collaboration, we used a reinforcement learning task known to reflect outcome-irrelevant learning, where irrelevant stimulus features (e.g., spatial location) are systematically assigned with value, resulting in suboptimal decisions. This design allowed us to test two competing hypotheses. The first posited that autistic individuals may engage more in value updating of irrelevant features, due to increased susceptibility to process irrelevant information. The second, based on an enhanced rationality framework, predicted that autistic individuals would show greater resistance to outcome-irrelevant learning and demonstrate more optimal learning compared to non-autistic individuals. Computational modeling and model-agnostic analyses revealed that participants reporting a clinical autism diagnosis, compared to non-autistic participants, exhibited substantially reduced outcome-irrelevant learning. Importantly, reduced outcome-irrelevant learning was also significantly correlated with higher autistic traits across the full population sample, suggesting that it reflects a broader, dimensional cognitive profile rather than only a categorical group difference. These findings demonstrate, for the first time, enhanced resistance to outcome-irrelevant learning biases in autism, revealing a domain in which autistic cognition may confer advantages for rational decision-making in noisy environments. Our results add to growing evidence for domain-specific cognitive advantages in autism and extend to suggest that this cognitive style, observed across the entire cohort, may offer a more adaptive approach to managing biases in decision-making.
Ben-Artzi et al. (Thu,) studied this question.