Children under age 6 or 7 quantify over objects of a single kind in ways that differ markedly from adults. For example, shown three pieces of a single fork and one whole fork, children say there are four forks, and not one or two forks. Here we explore why children exhibit these phenomena as well as motivating hypotheses of what's changing over ages 5 through 8. We note that partitive language such as "two pieces of a fork" specifies atoms of quantification and provides a context in which a piece of a fork is not a fork. We hypothesize that while children begin learning partitive language early in the preschool years, they assign it a precursor meaning that lacks adult quantificational force. Five experiments provide evidence that children begin learning partitive terms at age 2, but analyze these terms and phrases with precursor meanings akin to subsective adjectives. Experiments 1-4 establish that knowledge of partitive vocabulary is unrelated to quantificational decisions on a part/whole quantification task at these ages. In contrast, this early knowledge of partitive language selectively predicts success at using the contrast between "half/ piece" and "whole" in a non-verbal visual working memory (VWM) task that does not involve quantificational ambiguity. Furthermore, the level of success on the half/whole VWM task is comparable to that on small/big VWM task, where "small/big" are indisputably subsective adjectives. These data are consistent with the precursor meaning hypothesis. Experiment 5 directly tests whether 4-year-old children can make sense of quantificational language like "this piece and this piece make one fork" and finds they cannot. These data suggest that adult analysis of the quantificational force of partitive language is not mastered until after age 4. They motivate further research that investigates whether, beginning at age 5, a semantic/syntactic change from a precursor meaning to adult meaning of partitive language may reflect and contribute to a wider change in conceptual quantificational resources.

A longstanding assumption in language acquisition is that comprehension precedes production. However, for some words, children produce them long before they master their adult meanings, suggesting that such words are “hard to learn.” We investigate this phenomenon for the relational terms “same” and “different.” Although pre-lexical infants represent sameness and difference early in life, it is unknown when children acquire the corresponding word meanings. Using corpus analyses of English-learning children aged 2–4 years, we show that “same” and “different” are frequent in caregiver speech and are produced by all middle class English learning children by age 2. Nevertheless, three comprehension experiments with over 200 high SES English learning children aged 2–5 years reveal that only a minority of 2- and 3-year-olds understand the relational meanings of these words, with reliable comprehension emerging at age 4. Moreover, children almost never comprehend one word without the other, suggesting that “same” and “different” are learned in close temporal proximity. We discuss possible underlying reasons for this particular extended acquisition trajectory, including the different formats of non-linguistic and linguistic representations of sameness and difference. These findings clarify what makes certain relational words hard to learn and illuminate how language acquisition can impact relational thinking.

Many preschoolers do not deploy logical concepts like OR or POSSIBLY in the 3- and 4-container tasks, even at age 4. Gautam et al. (2021) report that with procedural changes, children with an age range of 2;6 to 3;0 perform well on the 4-container task, suggesting that performance demands of earlier implementations might have masked children’s logical competence. Here we replicate that result, but show it holds only on the first 2 trials of the task, after which performance is the same as all other implementations in the literature. Moreover, older 2-year-olds fail on the simpler 3-container task even with those procedural changes. Children’s success in Gautam et al.’s 4-container trials does not reflect the deployment of OR or POSSIBLY.

Complex sequences are ubiquitous in human mental life, structuring representations within many different cognitive domains-natural language, music, mathematics, and logic, to name a few. However, the representational and computational machinery used to learn abstract grammars and process complex sequences is unknown. Here, we used an artificial grammar learning task to study how adults abstract center-embedded and cross-serial grammars that generalize beyond the level of embedding of the training sequences. We tested untrained generalizations to longer sequence lengths and used error patterns, item-to-item response times, and a Bayesian mixture model to test two possible memory architectures that might underlie the sequence representations of each grammar: stacks and queues. We find that adults learned both grammars, that the cross-serial grammar was easier to learn and produce than the matched center-embedded grammar, and that item-to-item touch times during sequence generation differed systematically between the two types of sequences. Contrary to widely held assumptions, we find no evidence that a stack architecture is used to generate center-embedded sequences in an indexed AnBn artificial grammar. Instead, the data and modeling converged on the conclusion that both center-embedded and cross-serial sequences are generated using a queue memory architecture. In this study, participants stored items in a first-in-first-out memory architecture and then accessed them via an iterative search over the stored list to generate the matched base pairs of center-embedded or cross-serial sequences.

Two related methods have been widely used to test animals' capacity to reason by exclusion: Call's (2004) 2-cups task, where subjects choose between two cups in which one reward was hidden and the empty cup is revealed; and Premack & Premack's task (1994) where one reward goes in each of the two cups and one reward is removed. In both cases, success is identifying the cup with the reward. It has been suggested that among corvids, the demands associated with caching may foster exclusion abilities by reinforcing the experience of relevant contraries (e.g., empty vs. full), or by supporting domain general capacities that support caching and exclusion inferences. However, Shaw et al. (2013) found that Eurasian jays, proficient cachers, failed a version of Call's 2-cups task. To further test the caching hypothesis, we amended the procedure using invisible displacement to hide rewards, removing aspects of Shaw et al.'s procedure that may have masked competence of Eurasian jays. We tested seven Eurasian jays, who showed spontaneous success on both Call's 1-reward/show-empty (80% correct) and Premack and Premack's 2-reward/remove-1 (85% correct). These results align with the hypothesis that corvids' exclusion reasoning is related to caching behaviour. ### Competing Interest Statement The authors have declared no competing interest.

The classic Michottean 'launching' event is consistent with a real-world Newtonian elastic collision. Previous research has shown that adult humans distinguish launching events that obey some of the physical constraints on Newtonian elastic collisions from events that do not do so early in visual processing, and that infants do so early in development (< 9 months of age). These include that in a launching event, the speed of the agent can be 3 times faster (or more) than that of the patient but the speed of the patient cannot be detectably greater than the speed of the agent. Experiment 1 shows that 7-8-month-old infants also distinguish canonical launching events from events in which the motion of the patient is rotated 90(degrees) from the trajectory of the motion of the agent (another outcome ruled out by the physics of elastic collisions). Violations of both the relative speed and the angle constraints create Michottean 'triggering' events, in which adults describe the motion of the patient as autonomous but not spontaneous, i.e., still initiated by contact with the causal agent. Experiments 2 and 3 begin to explore whether infants of this age construe Michottean triggering events as causal. We find that infants of this age are not sensitive to a reversal of the agent and patient in triggering events, thus failing to exhibit one of the signatures of representing an event as causal. We argue that there are likely several independent events schemas with causal content represented by young infants, and the literature on the origins of causal cognition in infancy would benefit from systematic investigations of event schemas other than launching events.

What Babies Know (WBK) argues that core knowledge has a unique place in cognitive architecture, between fully perceptual and fully conceptual systems of representation. Here I argue that WBK's core knowledge is on the perception side of the perception/cognition divide. I discuss some implications of this conclusion for the roles language learning might play in transcending core knowledge.

Some nonlinguistic systems of representation display some of the six features of a language-of-thought (LoT) delineated by Quilty-Dunn et al. But they conjecture something stronger: That all six features cooccur homeostatically in nonlinguistic thought. Here I argue that there is no good evidence for nonlinguistic deductive reasoning involving the disjunctive syllogism. Animals and prelinguistic children probably do not make logical inferences.

Relational reasoning is a cornerstone of human cognition. Extensive work, drawing on the Relational Match to Sample paradigm (RMTS), has established that humans, at least above the age of five, are much more proficient relational reasoners than younger children or non-human animals. While sometimes differences between populations derive from differences in capacity (the capacity to create representations in a certain format or of a certain complexity, information processing capacity), other times such differences derive from different learning histories alone. Here we distinguish between two types of learning history explanations on the example of four-year-olds' failure on Premack's (1983) RMTS task: (1) that children four-year-olds have not yet created representations of the relations same and different with the properties need to support success on RMTS and (2) that four-year-olds have different inductive biases than do adults. Experiment 1 established that four-year-olds are at chance on the RMTS task we deploy as a transfer task in Experiment 2. Experiments 2A-C each provide children with a mere 8 trials of training on of one three MTS tasks (Number, Size and Identity MTS, respectively), none of which involves making matches of same to same or different to different. The very brief training (eight trials) on two of these tasks (Number MTS, Size MTS) leads to spontaneous success on RMTS in four-year-olds. Identity MTS has no effect on subsequent performance on RMTS. Given the brevity and non-relational nature of the training the successes after Number and Size MTS training must have resulted from changing inductive biases alone. Furthermore, the same two training tasks increased relational responding by adults on a related task (Kroupin & Carey, in press), whereas Identity MTS training did not, suggesting that the mechanisms through which the training changed inductive biases are at least partially continuous between ages four and adulthood.

Past work has found that infants show more interest when an object that has at least two properties of animate beings, such as engaging in self-generated motion and having fur, is shown to be hollow than when an object with none or one of these properties is revealed to be hollow. When an object is grabbed by a hand and moved to a new place, by 7 months of age, infants explain the motion of the object as due to the hand, and thus do not interpret this object as capable of self-generated motion. This constant application of force is called an "entraining" event. Other work has found that 6-month-old infants are sensitive to the reversals of causal roles in "launching" events (billiard-ball-like collisions), but not entraining events. Here, we examine whether 10-month-old infants explain the motion of the patient in a launching event as being due to the contact with the launching agent. Experiment 1 replicates past work, showing that infants look longer when a self-propelled object with animate features (fur or feathers) is shown to be hollow, compared to a similar object undergoing spatiotemporally identical motion entrained by a human agent. Experiment 2 finds that infants look equally at the agent and patient, both covered by fur or feathers, of a launching event when each is revealed to be hollow. Experiment 3 shows that infants look longer when a fur-covered causal patient is shown to be hollow compared to a plain-box causal agent, indicating that 10-month-old infants do not explain the motion of the causal patient of a launching event as due to the agent, even though they do so for an entraining event. This dissociation suggests the existence of multiple independent causal representations in the first year of life.

How do humans develop the capacity to reason? In five studies, we examined infants' emerging ability to make exclusion inferences using negation, as in the disjunctive syllogism (P or Q; not P; therefore Q). Inspired by studies of non-human animals and older children, Experiments 1-3 used an exclusion task adapted from Call's (2004) 2-cup paradigm and Experiments 4-5 used an exclusion task adapted from the blicket detector paradigm (Sobel & Kirkham, 2006). In both tasks, we found failure to make exclusion inferences at 15 months, fragile success at 17 months, and robust success by 20 months of age. These data converge with some prior evidence that fails to find a capacity to represent negation in infants younger than 15 months of age and conflict with other evidence from different paradigms that suggests infants do have this capacity. We discuss three different resolutions of these conflicting data, and suggest lines of further work that might adjudicate among them.

Nonhuman animals and preschoolers struggle with Relational-Match-to-Sample (RMTS), a classic test of the capacity for second-order relational, analogical, and reasoning. These failures are often explained by limitations in representational or computational capacities. Drawing on recent evidence for robust spontaneous RMTS success (i.e., without error-feedback) in crows and parrots after minimal second-order training, we present five experiments with human adults consistent with the possibility that population differences sometimes instead derive from differences in inductive biases alone. Experiment 1 confirms human adults have the capacities and requisite representations to succeed spontaneously on RMTS. Experiments 2-5 utilize a modified RMTS task in which adults make relational matches only about half of the time. Experiment 3 tests whether eight trials of various MTS tasks, nonsecond-order training featured in the aforementioned comparative studies, can increase spontaneous second-order relational responding in human adults. Two of the MTS tasks (Number, Size MTS) do so, demonstrating that MTS training can, in fact, increase relational responding by changing inductive biases alone. The other MTS tasks (Identity, Color MTS) do not do so, evidence that the facilitating effect is not a result of matching involved in MTS per se. Experiments 4 and 5 test one hypothesized mechanism by which specifically Number/Size MTS tasks may have led to increased relational responding, that is, by inhibiting preexisting biases to match on shape and/or color, making relational matches relatively more likely. We close by discussing the importance of research into inductive biases to the project of understanding relational reasoning. (PsycInfo Database Record (c) 2022 APA, all rights reserved).

Young children do not always consider alternative possibilities when planning. Suppose a prize is hidden in a single occluded container and another prize is hidden in an occluded pair. If given a chance to choose one container and receive its contents, choosing the singleton maximizes expected reward because each member of the pair might be empty. Yet, 3-y-olds choose a member of the pair almost half the time. Why don’t they maximize expected reward? Three studies provide evidence that 3-y-olds do not deploy possibility concepts like MIGHT, which would let them represent that each container in the pair might and might not contain a prize. Rather, they build an overly specific model of the situation that correctly specifies that the singleton holds a prize while inappropriately specifying which member of the pair holds a prize and which is empty. So, when asked to choose a container, they see two equally good options. This predicts approximately 50% choice of the singleton, observed in studies 1 and 3. But when asked to throw away a container so that they can receive the remaining contents (study 2), they mostly throw away a member of the pair. The full pattern of data is expected if children construct overly specific models. We discuss whether 3-year-olds lack possibility concepts or whether performance demands prevent deployment of them in our tasks.

My work in cognitive science has given me immeasurable pleasure for over 60 years. Here I trace how I have come to my current understanding of conceptual development. I emphasize the roles of accident and luck along my path as well as the importance of being able to deal with failures. I also place my career in context of the rest of my life.

A pendulum swing in views of cognitive development has occurred over the past 20 years. The issue in the pendulum swing is the proper description of the child's cognitive capabilities – the proper description of how the child's conceptual system differs from the adult's. A commitment to stages is a commitment to there being a domain-general, i.e., conceptual content free, level of description of what changes in the course of cognitive development. The pendulum has swung away from a commitment to stages in the Piagetian sense. Essences are the ultimate reality, the real nature of the things in nature, and are usually deeply hidden. It is the business of science both to discover essences and to discover deeper realities underlying surface appearances. The discovery of the essences of the things in nature and the discovery of deeper realities underlying surface phenomena are two of the goals of theory development.

How do humans develop the capacity to reason? In five studies, we examined infants’ emerging ability to make exclusion inferences using negation, as in the disjunctive syllogism (P or Q; not P; therefore Q). Inspired by studies of non-human animals and older children, Experiments 1-3 used an exclusion task adapted from Call’s (2004) 2-cup paradigm and Experiments 4-5 used an exclusion task adapted from the blicket detector paradigm (Sobel & Kirkham, 2006). In both tasks, we found failure to make exclusion inferences at 15 months, fragile success at 17 months, and robust success by 20 months of age. These data converge with some prior evidence that fails to find a capacity to represent negation in infants younger than 15 months of age and conflict with other evidence from different paradigms that suggests infants do have this capacity. We discuss three different resolutions of these conflicting data, and suggest lines of further work that might adjudicate among them.

Kind representations, concepts like table, triangle, dog, and planet, underlie generic language. Here, we investigate the formal structure of kind representations-the structure that distinguishes kind representations from other types of representations. The present studies confirm that participants distinguish generic-supporting properties of individuals (e.g., this watch is made of steel) and accidental properties (e.g., this watch is on the nightstand). Furthermore, work dating back to Aristotle establishes that only some generic-supporting properties bear a principled connection to the kind, that is, are true of an individual by virtue of its being a member of a specific kind (e.g., telling time for a watch). The present studies tested the hypothesis that principled connections are part of the formal structure of kind representations. Specifically, they tested whether they structure a newly learned kind representation. Experiment 1 found that introducing a property of a newly encountered novel kind in any one of four linguistic frames that provide evidence that a property has a principled connection to a kind (e.g., "It has fur because it is a blick") led participants to infer a different conceptual consequence of principled connections (i.e., "There is something wrong with this blick, which does not have fur") for which they had no direct evidence. Two introduction frames that provided no evidence for principled connections (e.g., "Almost all blicks have fur") did not generate the same consequence. Experiment 2 found that all of the targeted properties were generic licensing, irrespective of the introduction frame. That the distinction between properties that bear principled connections to their kinds, and merely generic-supporting properties structures novel kind representations, provides strong evidence that this distinction is part of the formal structure of kind representations.

A much replicated finding is that only humans above the age of five succeed spontaneously on Premack’s Relational Match to Sample task, which has been widely used in the comparative and developmental literatures to probe relational reasoning capacities. We review four different types of explanations for the failures of young children and non-human animals, two that posit capacity limitations and two that posit differences in learning histories alone. We review training studies that rule out capacity limitations, at least for crows, parrots, four-year-old children, and a variety of primate species. Finally, we review recent studies demonstrating that population differences sometimes reflect differences in inductive biases alone and discuss the crucial importance of inductive biases in relational reasoning.