Humans we manipulate objects in innovative ways to navigate our environment and achieve goals. Our study investigates the ability to discover new object functions (a form of tool innovation) in human children and adults and in large language models such as GPT and FLAN. We introduced 42 three- to seven-year-olds and 40 adults to problem situations in which the typical tools were missing. Participants were asked to choose among objects that were superficially related in terms of taxonomy and perceptual features but functionally useless, objects that were superficially dissimilar but functionally relevant, and objects that were neither superficially nor functionally relevant. Both preschoolers and adults selected functionally appropriate tools, despite the lack of a direct relationship between the functionally appropriate tools and the standard tools used to achieve the relevant goals. In contrast, large language models, which are trained on unprecedentedly immense amounts of text, did not near human-level performance.

Three-dimensional objects pose a challenge for our visual system, since we can only view objects from a single limited perspective at a given moment. Previous work found that given a limited perspective, infants represent 3D objects as complete volumes. Our study replicated this finding in 4- to 7-year-olds and adults, using an explicit prediction measure rather than looking times. We also explored whether humans have a bias to represent visually limited 3D objects as symmetrical rather than asymmetrical across shape, size, texture, and color. Overall, there was an above-chance preference for full volumetric and symmetrical object completion that increased with age. Low-level perceptual similarity of choices did not predict participants’ choices. Moreover, we evaluated ResNet-50 neural networks on the same tasks: they represented objects as complete volumes, but did not show substantial preference for symmetrical 3D representations. This raises the possibility that incorporating human symmetry biases could improve computer vision.

Autonomous agents often need to decide between choosing actions that are familiar and have previously yielded positive results (exploitation) and seeking new information that could help uncover more effective actions (exploration). We present an “observe or bet” task that separates “pure exploration” from “pure exploitation”: 75 five-to-seven-year-old children, 60 adults and computational agents have to decide either to observe an outcome without reward, or to bet on an action without immediate feedback at varying probability levels. Their performances were measured against solutions from the partially observable Markov decision process and meta-RL models. Children and adults tended to choose observation more than both algorithm classes would suggest. Children also modulated their betting policy based on the probability structure and amount of evidence, exhibiting “hedging behavior” a strategy not evident in standard bandit tasks. The results provide a benchmark for reasoning about reward and information in humans and neural network models.