Title: Hypothesis Generation and Inductive Inference in Children and Language Models

Abstract:

Navigating real-world decision-making necessitates the construction of mental models amidst uncertainty regarding evidence, underlying causal mechanisms, and the state of the world. This study investigates which computational principles drive human inference under such conditions and whether large language model (LLM)-based agents display comparable behaviors when subjected to identical constraints. To answer these questions, we employed an inductive inference Box Task, wherein both human children and LLM-based agents attempted to deduce a latent cause through sequential interactions with an uncertain environment. We formalized this task as program induction utilizing Bayesian particle-based inference, offering two distinct yet complementary perspectives: first, as a constraint satisfaction process involving hypotheses, and second, as a program synthesis challenge where hypotheses function as executable programs assessed against evidence.

Through the lens of constraint-based formulation, our analysis reveals that children’s actions are most accurately described by a blend of subjective evidence reliability and online hypothesis generation. This approach successfully accounts for their evidence-seeking habits as well as the observed disconnect between task completion and the generalization of rules. Conversely, adopting the program synthesis framework allows us to utilize LLM-based agents as model organisms—controllable systems enabling the systematic manipulation of task parameters. Across various backends, these agents mirrored children’s reactions to variations in evidence reliability and observability. Specifically, they demonstrated tendencies to discount unreliable data, seek clarification on partial information, and exhibit a similar dissociation between completing the task and generalizing causal rules. However, LLM-based agents also displayed a propensity to over-observe and over-comply with instructions compared to their human counterparts. These findings indicate that while children and LLM-based agents adapt in similar ways to environmental structures, their information-seeking behaviors are governed by distinct inductive biases and underlying costs.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC