Hypothesis Generation and Inductive Inference in Children and Language Models

Large language model agents and children adapt similarly to environmental structure during inductive reasoning tasks, but their information-seeking behaviors diverge, according to a study posted to arXiv on May 23, 2026 [1]. The research used an inductive inference Box Task in which both human children and LLM-based agents inferred a latent cause through sequential interaction with an uncertain environment [1]. The task was formalized as program induction with Bayesian particle-based inference, providing two complementary interpretations: a constraint satisfaction process over hypotheses, and a program synthesis problem where hypotheses are executable programs evaluated against evidence [2]. Children's behavior was best explained by a combination of subjective evidence reliability and online hypothesis generation, which accounted for both their evidence-seeking patterns and a dissociation between task completion and rule generalization [1]. The study treated LLM-based agents as model organisms — controllable systems that allow systematic manipulation of task conditions [2]. Across multiple backends, the LLM-based agents replicated children's responses to changes in evidence reliability and observability. They discounted unreliable evidence, sought to resolve partial information, and dissociated between task completion and causal generalization [1]. However, the agents tended to over-observe and over-comply with instructions relative to the children [2]. The findings contribute to a broader effort to understand whether computational systems can replicate human-like inference. Artificial intelligence research has long drawn on psychology, linguistics, and philosophy to model reasoning and learning [3]. The concept of abductive reasoning — inferring the most likely explanation from incomplete observations — was formally defined by the American philosopher and logician Charles Sanders Peirce in the 19th century [4]. Peirce's work on logic and semiotics foreshadowed debates that dominated 20th-century Western philosophy and later informed computational approaches to inference [4]. Developmental psychology has similarly probed how children construct causal models from sparse data. Researchers such as Fei Xu, a professor of psychology at UC Berkeley who directs the Berkeley Early Learning Lab, have focused on cognitive and language development from infancy through middle childhood [5]. The new study extends this tradition by placing child and machine behavior side by side under matched constraints [1]. The authors suggest that while children and LLM-based agents adapt similarly to environmental structure, their information-seeking behavior exhibits distinct underlying costs and inductive biases [2]. The preprint has not yet been peer-reviewed.