Decades of research on infant physical reasoning have revealed that the ability to reason about objects’ causal interactions in simple physical events develops gradually over the first two years of life (for a review, see Lin et al., 2022). At least two factors contribute to this protracted development. The first is that infants must identify, for each event category (e.g., occlusion, containment, support, and collision events), what features of objects and their arrangements are causally relevant for predicting outcomes. With respect to occlusion features, for example, infants typically identify height and width—or size more generally—by 3.5 months (Baillargeon and DeVos, 1991, Wang et al., 2004), shape by 4.5 months (Wilcox, 1999, Wilcox and Baillargeon, 1998b), pattern by 7.5 months (Wilcox, 1999), and color by 11.5 months (Lin et al., 2021, Wilcox, 1999). The evidence that infants have identified these occlusion features comes from a variety of tasks. In the case of height and width information, for example, infants ages 3.5–6 months detect an interaction violation if a tall object becomes hidden behind a short occluder (Baillargeon and DeVos, 1991, Hespos and Baillargeon, 2001, Mou and Luo, 2017) or if a wide object becomes hidden behind a narrow occluder (Wang et al., 2004); they detect a change violation if a large ball changes into a small one when passing behind an occluder too narrow to hide both balls (Wilcox, 1999); they detect a change violation if a tall object changes into a short one, or if a wide object changes into a narrow one, when briefly held behind a large occluder (Goldman & Wang, 2019); and they reach preferentially for a tall as opposed to a short occluder when searching for a tall object (Hespos & Baillargeon, 2006).

The second factor that contributes to the protracted development of early physical reasoning is that infants who have identified features as relevant to an event category nevertheless fail to use these same features to individuate objects in events from the category (Xu and Carey, 1996, Xu et al., 2004). For example, when 12-month-olds see two objects that differ in size (e.g., a small and a large ball, both red and covered with glitter), in color (e.g., a ball with pink and green stripes and a ball with purple and orange stripes), or in size, pattern, and color (e.g., a small soccer ball with orange, green, and white hexagons and a large red ball covered with glitter) emerge in alternation from behind a large occluder, they fail to detect an individuation violation when the occluder is finally removed to reveal only one of the objects (Xu et al., 2004).

In this article, we focus on this second factor. Why do infants who have identified features as causally relevant for occlusion events fail to use these features when individuating objects? If infants detect a change violation when a single object surreptitiously changes size, pattern, or color behind an occluder (Goldman and Wang, 2019, Lin et al., 2021, Wang and Baillargeon, 2006, Wilcox, 1999), why do they fail to establish separate representations when two objects that differ in size, pattern, or color emerge in alternation from behind an occluder?

Ever since the discovery of infants’ baffling difficulty at individuating objects, researchers have been seeking to understand its cause. Over the course of these investigations, additional findings have come to light that have helped inform the quest for an explanation. Our article had two main goals. The first was to review key findings from standard and modified individuation tasks and to discuss three accounts that have been proposed for these findings: the kind account, the mapping account, and the more recent two-system account. As will become clear, the two-system account borrows heavily from the earlier accounts but still differs from them in critical respects. In particular, although the kind and two-system accounts agree that categorical information plays an essential role in infants’ ability to individuate objects, they differ widely in their claims about why this information matters and, relatedly, about what types of categorical information can support infant individuation. The second goal of our article was to examine these competing claims experimentally, using two novel manipulations.

In this section, we summarize findings on infant individuation from standard tasks and from two modified tasks. These findings have brought to light both strengths and limitations in infants’ ability to individuate objects.

In standard tasks, infants see a sequence of two events (see Fig. 1). In the first event, two different objects are brought out in alternation from a hiding location; the second event varies across tasks but is intended to assess whether infants correctly individuated the two objects in the first event. In some tasks, infants see an occlusion event followed by a no-occlusion event. The two objects first emerge in alternation on either side of a large screen, which is then lowered to reveal only one of the objects; surprise at this violation is taken to indicate that infants correctly inferred how many objects were present behind the screen (we refer to these tasks as standard-violation tasks; Wilcox and Baillargeon, 1998a, Xu and Carey, 1996). In other tasks, infants see a containment event followed by a search event. The two objects are first lifted in alternation from inside a large box, which is then moved within infants’ reach; persistent searching after one object has been retrieved is taken as evidence that infants correctly inferred how many objects were present inside the box (we refer to these tasks as standard-search tasks; Van de Walle et al., 2000).

A wealth of research indicates that infants succeed at standard tasks only if they are able to assign the two objects introduced in the first event to distinct categories. Because there is substantial development with age in what categorical information infants spontaneously encode about objects, whether they succeed at a task depends on (a) whether the two objects are categorically distinct in some way and (b) whether this categorical distinction is one that infants happen to encode at the age they are tested.

Prior to their first birthday, most infants do not spontaneously encode an isolated object’s basic-level category, such as toy duck, ball, or block (Pauen, 2002, Xu and Carey, 1996). However, they do encode broad, ontological categorical information, such as whether the object is human or non-human (Bonatti et al., 2002, Kibbe and Leslie, 2019), animate or inanimate (Setoh et al., 2013, Surian and Caldi, 2010), and self-propelled or inert (Decarli et al., 2020, Luo et al., 2009). Thus, 9–10-month-olds succeed at standard tasks when tested with a humanlike vs. a non-humanlike object (e.g., a doll vs. a toy dog; Bonatti et al., 2002; see also Bonatti et al., 2005, Stavans et al., 2019, Stavans and Baillargeon, 2019), an animate vs. an inanimate object (e.g., a flying bee vs. a block carried by a hand; Surian & Caldi, 2010), or a self-propelled vs. an inert object (e.g., a self-propelled ball vs. an identical ball carried by a hand; Decarli et al., 2020).

Around their first birthday, infants begin to spontaneously encode objects’ basic-level categories (Pauen, 2002, Xu and Carey, 1996). As a result, infants now also succeed at standard tasks when tested with two objects from different basic-level categories such as a toy duck vs. a ball (Xu & Carey, 1996), a ball vs. a block (Wilcox & Baillargeon, 1998a), or a toy telephone vs. a toy book (Van de Walle et al., 2000).

Critically, the main point from these investigations is not that infants succeed at standard tasks when they can encode categorical information about the two objects shown in the first event, and fail otherwise: As the preceding descriptions indicate, infants always encode categorical information about the objects (e.g., a ball and a block might both receive categorical descriptors such as non-humanlike, inanimate, rigid, and closed). Rather, the main point is that infants succeed when they can assign one or more distinct categorical descriptors to the two objects, and fail otherwise. Thus, prior to 12 months, infants fail if tested with two objects they encode as belonging to the same ontological categories and differing only in their featural properties (e.g., a toy duck vs. a ball; Xu & Carey, 1996; see also Bonatti et al., 2002, Stavans et al., 2019, Stavans and Baillargeon, 2019, Van de Walle et al., 2000, Wilcox and Baillargeon, 1998a). Similarly, at 12 months, infants fail if tested with two objects they encode as belonging to the same ontological and basic-level categories and differing only in their featural properties (e.g., a small yellow cup decorated with green and red stripes and fitted with an orange handle vs. a large semi-transparent cup decorated with multicolored shapes and fitted with a blue handle; Xu et al., 2004; see also Bonatti et al., 2005).

Although prior to 12 months infants do not yet encode objects’ basic-level categories, they can be induced to do so via various experimental manipulations, with positive effects on their individuation performance (Futó et al., 2010, Stavans and Baillargeon, 2018, Xu, 2002). For example, in standard-violation tasks, 9-month-olds succeeded with a lexical manipulation: As each object emerged into view in the occlusion event, an experimenter gave the object a distinct label (e.g., “Look, [baby’s name], a duck!”, “Look, [baby’s name], a ball!”; Xu, 2002). Similarly, 10-month-olds succeeded with a pedagogical-functional manipulation: As each object was brought out in the occlusion event, an experimenter produced ostensive-communicative signals (“Hi baby, hi!”) and demonstrated the object’s distinct function (e.g., a radio that played a melody when its dial was turned, a lamp that flashed small lights when its handle was pulled; Futó et al., 2010). Finally, even 4-month-olds succeeded with a prior functional manipulation: In two introduction trials, an experimenter demonstrated two simple tools’ distinct functions (e.g., a masher that was used to compress sponges, tongs that were used to pick them up); next, in a test trial, the two tools emerged in alternation from behind a screen, which was then lowered to reveal only one of the tools (Stavans & Baillargeon, 2018).

The preceding findings make clear that contrastive categorical information, whether encoded spontaneously or via experimental manipulations, enables infants to succeed at standard tasks. Strikingly, when such information is unavailable, infants fail at standard-violation tasks not only when the screen is lowered to reveal only one of the objects, but also when it is lowered to reveal no objects at all. Thus, after seeing two red cups that differed only in pattern (yellow stripes vs. green dots) emerge in alternation from behind a screen, 12-month-olds failed to detect a violation when the screen was lowered to reveal either one cup or no cup. If the same cup appeared on either side of the screen, however, infants expected to see one cup and were surprised to see no cup, making clear that they were not simply confused by the no-cup outcome (Stavans et al., 2019, Stavans and Baillargeon, 2019). Similar results were found in a novel standard-violation task involving a containment event followed by a shaking event. To start, 9-month-olds saw a toy wolf and a ball being lifted in alternation from inside a large box. Next, the box was shaken vigorously yet produced no impact sounds at all, as though empty. Infants failed to detect a violation in this outcome, though they did expect to hear impact sounds if they saw both objects simultaneously, or a single object, at the start of the event (Stavans et al., 2019, Stavans and Baillargeon, 2019).

Together, the results reviewed in this section point to three conclusions. First, infants succeed at a standard task when they can assign distinct categorical descriptors to the two objects shown in the first event. Second, when they are unable to do so and encode the two objects as merely featurally distinct, they fail at the task. Finally, unsuccessful infants are not merely uncertain about whether one or two objects should be present in the second event: Rather, they hold no expectation at all about the event, resulting in a catastrophic failure.

We have just seen that infants fail at standard tasks when they encode the two objects shown in the first event as categorically similar and differing only in their featural properties. Are there circumstances in which infants demonstrate successful individuation even though they have to rely on the objects’ featural differences to do so? Modified tasks such as one-event tasks (described in this section) and remainder tasks (described in the next section) provide a positive answer to this question (see Fig. 1).

In one-event tasks, infants see a single, ongoing occlusion event involving two objects, and their ability to individuate the objects is assessed during the event itself. In some tasks, a transparent panel stands behind the screen and is revealed when the screen is lowered; infants thus see a continuing occlusion event involving first an opaque and then a transparent occluder. Wilcox and Chapa, 2002, Wilcox and Schweinle, 2002 found that under these conditions, 9.5-month-olds succeeded in individuating a ball and a block (recall that at this age infants encode such objects as categorically similar and differing only in their featural properties). After seeing the two objects emerge in alternation from behind a large screen, infants detected a violation if the screen was lowered to reveal only the ball visible behind the transparent panel; this effect was eliminated, as usual, when the panel was absent. Stavans et al., 2019, Stavans and Baillargeon, 2019 replicated this finding in a one-event task testing 10-month-olds with two different female dolls: One was a light-skinned, blue-eyed blonde whose long hair was decorated with blue streaks and worn in two braids, and the other was a dark-skinned, brown-eyed brunette whose long, loose hair was decorated with a bow (there were thus multiple featural differences, including size, shape, and pattern differences, that infants could use to distinguish the dolls). After seeing the two dolls emerge in alternation from behind a large screen, infants detected a violation if the screen was lowered to reveal only one of the dolls behind the transparent panel; this effect was eliminated when no panel was present.

Other one-event tasks come from neuroimaging studies conducted by Wilcox and her colleagues using functional near-infrared spectroscopy (Wilcox et al., 2012, Wilcox et al., 2014). In these tasks, objects appeared in alternation on either side of a large screen: a green ball and a green block (different-shapes event), a green ball and a red ball (different-colors event), or the same green ball (same-object event). Each event was repeated twice, and then the trial ended; the screen was never lowered, so that infants saw a single, ongoing occlusion event. During the event, activation was recorded in infants’ anterior temporal cortex, a brain region previously shown to be involved in the individuation process (Wilcox et al., 2010). Relative to the same-object event, 11.5-month-olds showed increased activation to both the different-shapes and different-colors events, whereas 5–7.5-month-olds showed increased activation to the different-shapes event only. Thus, consistent with prior findings (reviewed earlier) on the typical ages at which different occlusion features are identified, the older infants included shape and color information in their event representations and inferred that two objects were involved in both the different-shapes and different-colors events. In contrast, the younger infants included only shape information and inferred the presence of two objects in the different-shapes event only.

The one-event tasks described above suggest the following conclusion. When infants see two objects they encode as categorically similar emerge in alternation from behind a screen, they can give evidence of successful individuation if (a) they have identified one or more of the features that distinguish the objects as causally relevant for occlusion events and (b) their ability to individuate the objects is assessed during the occlusion event itself.

Remainder tasks provide further evidence that under certain circumstances infants can demonstrate successful individuation even when tested with objects they encode as differing only in their featural properties. These tasks are similar to standard tasks with one exception: At the end of the first event, only one object remains in the hiding location, because the other object is in a different location.

In some tasks, only one object is in the hiding location at the end of the first event because the other object has been left out in plain view. For example, in studies by Wilcox and her colleagues (Wilcox and Baillargeon, 1998a, Wilcox and Schweinle, 2002, Wilcox and Chapa, 2002), 5.5–9-month-olds first saw a block move along a platform and disappear behind a large screen. Next, a ball emerged on the other side of the screen and paused in full view on the platform. The screen was then lowered to reveal no block—only the ball could be seen to the right of the screen. Infants detected the violation in this outcome, suggesting that (a) they correctly individuated the block and the ball based on the featural information available, (b) they realized that the block was still behind the screen, and (c) they expected to see the block when the screen was lowered. In another study in this series (McCurry et al., 2009), the screen was made of cloth fringe; after the ball emerged from behind the screen and paused, the platform and screen were moved within infants’ reach. Infants reached significantly more for the screen than for the ball, suggesting that they believed the block was behind the screen and they wanted to retrieve it by reaching through the fringed screen. In both studies, the effect was eliminated when just the ball appeared on either side of the screen.

The preceding studies all used a very brief occlusion event, with no reversal in the movement path of either object, but similar results have been found with longer occlusion events. In one study with 11-month-olds, for example, two blocks that differed only in pattern were brought out three times in alternation from behind a large screen, and the last block to be brought out was left in plain view to the left of the screen (Lin & Baillargeon, 2019). Infants detected a violation when the screen was lowered to reveal no block, so that the only block visible was the one next to the screen. This effect was eliminated, as usual, if both blocks were returned behind the screen before it was lowered.

In other remainder tasks, only one object is in the hiding location at the end of the first event because the other object has been visibly moved out of the apparatus. For example, in a remainder study by Stavans et al., 2019, Stavans and Baillargeon, 2019 adapted from their novel standard-violation task described earlier, 9-month-olds first saw a toy wolf and a ball being lifted in alternation from inside a large box. Next, the last object lifted was removed from the apparatus, so that only the other object remained inside the box. When the box was next shaken, infants expected impact sounds, suggesting that (a) they correctly individuated the two objects based on the featural information available, (b) they realized that one object was still inside the box, and (c) they expected this object to collide noisily with the rigid walls of the box during the shaking.

The remainder tasks described above suggest the following conclusion. When infants see two objects they encode as categorically similar emerge in alternation from a hiding location, and this first event is followed by a second event designed to assess their ability to individuate the objects, infants succeed at the task if (a) they have identified one or more of the features that distinguish the objects as causally relevant for the event category depicted in the first event (e.g., occlusion, containment) and (b) only one object remains in the hiding location at the end of the first event.1

Over the past three decades, several accounts have been offered for the findings reviewed in the previous section. In this section, we focus on two long-standing accounts, the kind and mapping accounts; in the next section, we introduce the more recent two-system account, which builds on these earlier accounts.

The kind and mapping accounts offer different explanations for infants’ failure at standard tasks when tested with two objects they encode as merely featurally distinct. The kind account portrays this failure as an individuation failure: Infants are unable to use the featural information available to determine how many objects are involved in the first event. In contrast, the mapping account portrays this failure as a mapping failure: Infants can use the featural information available to establish separate representations for the two objects in the first event, but they have difficulty using these feature-based representations when interpreting the second event.

The kind account holds three main assumptions (Carey, 2009, Xu, 2002, Xu, 2007, Xu, 2023, Xu and Baker, 2005, Xu and Carey, 1996, Xu and Carey, 2000, Xu et al., 2004). First, infants succeed at standard tasks when they can assign the two objects to distinct kinds (and more specifically sortals, the subset of kind concepts having to do with object concepts). Kinds are stable, long-term object categories, often with causally or functionally related features, that support individuation and identification; members of a kind share a number of intrinsic, stable properties and cannot spontaneously change into members of a different kind. This means that if two objects emerge in alternation from behind a screen and infants represent each object as belonging to a different kind, they can infer the presence of two objects because they understand that objects cannot spontaneously change kinds.

Second, infants are initially able to represent only a few kinds, such as inanimate objects, self-propelled objects, animate entities, and humanlike entities (some of these findings were initially taken to support the object-first hypothesis, the human-first hypothesis, and so on; Bonatti et al., 2002, Decarli et al., 2020, Surian and Caldi, 2010, Xu and Carey, 1996). By their first birthday, however, infants can also represent basic-level kinds such as ball, cup, and toy duck. At least three processes are thought to contribute to infants’ acquisition of basic-level kinds: (a) as they interact with objects, infants gradually learn for each object kind what featural properties remain constant over time; (b) as they hear object labels, infants innately assume that distinct labels refer to distinct object kinds; and (c) as they observe objects’ functional demonstrations, infants innately assume that distinct functions specify distinct object kinds.

Finally, although contrastive kind information is necessary for success at standard tasks, contrastive featural information may be sufficient for success at modified tasks with lower information-processing demands, such as one-event and remainder tasks. As we saw in the last section, these tasks have one event rather than two (one-event tasks), or they have one object rather than two in the hiding location at the end of the first event (remainder tasks). These reduced processing demands allow infants to correctly individuate objects based on their featural differences alone.

Although the kind account provides a possible explanation for many of the findings we have discussed, it leaves several questions unanswered. First, consider once again the findings of Xu et al. (2004). If by 12 months most infants have acquired the basic-level kind ‘ball’ and understand that balls do not spontaneously change featural properties, why do they fail at standard tasks involving two balls that differ in size, color, and pattern? Second, the kind account refers to standard tasks as “is-it-one-or-two” tasks (Xu, 2007) and assumes that unsuccessful infants are uncertain whether the hiding location holds a single object with changing properties or two different objects with stable properties. As we have seen, however, unsuccessful infants are not surprised even when given evidence that there are no objects at all in the hiding location. Finally, it is unclear why contrastive kind information is necessary for success at standard tasks, whereas contrastive featural information is sufficient for modified tasks such as one-event and remainder tasks. After all, the differences between standard and modified tasks can be rather subtle. For example, in all of these tasks, infants might see two objects emerge the same number of times from behind a screen; the tasks might differ only in that (a) a transparent panel is revealed when the screen is lowered (one-event tasks) or (b) the last object to emerge is left out in view before the screen is lowered (remainder tasks). Why do these task modifications allow infants, from a young age, to individuate objects they encode as merely featurally distinct? What mechanisms might be at play?

The mapping account holds three main assumptions (Needham and Baillargeon, 2000, Wilcox, 2003, Wilcox and Baillargeon, 1998a, Wilcox and Chapa, 2002, Wilcox and Schweinle, 2002, Wilcox and Schweinle, 2002, Wilcox and Chapa, 2002, Wilcox et al., 2003, Wilcox and Schweinle, 2003). First, when tested with a standard task, infants correctly individuate the two objects in the first event as long as these differ (a) in their categorical descriptors or (b) in one or more features infants have identified as causally relevant for the event’s category. This is because infants’ physical reasoning is assumed to be innately constrained by a principle of persistence, which states that all other things being equal, objects persist in time and space with all of their properties (Baillargeon, 2008, Lin et al., 2022). Thus, if two objects that differ in size, pattern, and color emerge in alternation from behind a screen, and infants have identified these features as relevant for occlusion events, they can infer that two objects are present because they realize that objects cannot spontaneously change categorical descriptors or featural properties.

Second, to succeed at a standard task, infants must not only be able to establish separate representations for the two objects in the first event: They must also be able to use these object representations when interpreting the second event. This is referred to as a mapping process: Infants must retrieve the object representations they established during the first event and map them onto the objects in the second event, to judge whether the two events are consistent. Importantly, whereas category-based object representations are easily mapped from the first onto the second event, feature-based object representations are not. It is assumed that categorical descriptors act as summary object representations. When retrieved for the mapping process, these summary representations take up fewer information-processing resources than do lists of featural properties and, as such, they make the process easier for infants to complete.

Third, in certain tasks, infants can demonstrate that they correctly individuated the two objects shown even when they had to rely on featural information to establish the objects’ representations. Thus, infants succeed at using feature-based object representations in one-event tasks because there is no mapping process required: Infants see a single, ongoing event, and they simply monitor its progress to assess whether it is consistent with their physical knowledge. Moreover, although remainder tasks, like standard tasks, do present infants with a sequence of two events and require mapping, the fact that only one object is in the hiding location at the end of the first event makes the mapping of feature-based object representations easier to complete.

Like the kind account, the mapping account provides an explanation for many of the findings we have reviewed but still leaves several questions unanswered. In particular, why are infants who fail at standard tasks not surprised even when given evidence that there are no objects at all in the hiding location? Shouldn’t the disappearance of two objects register as a persistence violation? Second, the mapping account offers little detail about the nature of the mapping process, the difficulties it causes, and the mechanisms by which categorical descriptors help infants overcome these difficulties. Finally, the mapping account also offers little information about the differences between mapping tasks with high processing demands, such as standard tasks, and mapping tasks with fewer processing demands, such as remainder tasks.

The two-system account incorporates two chief notions from the accounts discussed in the previous section. One notion, borrowed from the kind account, is that contrastive categorical information is critical for success at standard tasks. The other notion, borrowed from the mapping account, is that particular processes come into play when an event comes to an end, so that infants respond differently when faced with a single, ongoing event as opposed to a sequence of events involving the same objects. As will become clear, however, the two-system account offers different explanations for these notions.

According to the two-system account, infants’ successes and failures at standard and modified individuation tasks cannot be understood without spelling out in some detail the cognitive architecture that underlies infants’ reasoning about physical events (Baillargeon et al., 2012, Levine and Baillargeon, 2016, Lin et al., 2021, Lin et al., 2022, Stavans et al., 2019, Stavans and Baillargeon, 2019). As such, the account builds on related work from the adult and infant literatures (Gordon and Irwin, 1996, Huttenlocher and Lourenco, 2007, Kahneman et al., 1992, Leslie et al., 1998, Pylyshyn, 2007, Rips et al., 2006, Simons and Levin, 1998, Zacks, 2010) and is summarized here in terms of four assumptions.

First, when infants begin to attend to a physical event, at least two cognitive systems become involved, the object-file (OF) system (Gordon and Irwin, 1996, Kahneman et al., 1992, Leslie et al., 1998) and the physical-reasoning (PR) system (Lin et al., 2021, Lin et al., 2022, Stavans et al., 2019, Stavans and Baillargeon, 2019). The OF system builds a temporary file for each attended object using incoming perceptual information as well as information stored in memory, and it updates the file as additional information becomes available. Each file contains identity (“what”) and spatiotemporal (“where”) information, and each type of information includes categorical descriptors and fine-grained features. The PR system brings to bear its physical knowledge to represent the causal interaction depicted in the event and to predict how it will unfold; this physical knowledge includes core principles (persistence, gravity, and inertia), core concepts, and learned rules identifying causally relevant features for each event category. During the event, the OF and PR systems exchange information in two main ways. To start, the OF system passes on to the PR system the identity and spatiotemporal categorical descriptors in its object files; the PR system then uses that information to identify the event’s category (e.g., an occlusion event) and to assign event-specific causal roles to the objects in the event (e.g., occluder, occludee). Next, the PR system taps the OF system for information about previously identified features for the event category selected (e.g., at 12 months, this would include information about the relative sizes of the occluder and occludee as well as information about the shape, pattern, and color of the occludee). This featural information is then added to the event’s representation, updated as it changes, and interpreted by the PR system using its physical knowledge.

Second, because the OF system uses categorical information to individuate objects whereas the PR system uses both categorical and featural information, disagreements can arise between the two systems. To illustrate, imagine that 12-month-olds see an occlusion event in which two balls that differ in size, pattern, and color emerge in alternation from behind a screen. The OF system will conclude, based on the identity and spatiotemporal categorical descriptors at its disposal, that a single object is involved in the event. In contrast, the PR system will conclude, based on the featural information it added to the event’s representation, that two distinct objects are involved in the event. As one-event tasks demonstrate, this disagreement will matter little during the event itself because the PR system will be in charge of guiding infants’ responses: Its interpretation (two objects) will take precedence over that of the OF system (one object), leading infants to respond correctly to the event.

Third, when an event ends, the OF and PR systems must agree on the number of objects present for the event’s representation to be placed in memory and available to guide infants’ interpretation of the next event. Returning to our occlusion event, if the OF system (one object) and the PR system (two objects) initially disagree about how many objects are present, but this disagreement can be reconciled as the event comes to a close, the event’s representation can still be stored in memory. As remainder tasks demonstrate, reconciliation is possible if the two objects occupy distinct locations before the screen is lowered. The OF system will signal that a single object is present, next to the screen, whereas the PR system will signal that in addition to that object, there is another object present, behind the screen. Because the two objects have distinct spatiotemporal categorical descriptors (e.g., next to vs. behind the screen), the OF system will be able to open a file for the second object. As the OF and PR systems now agree that two objects are present, the event’s representation will be stored in memory, enabling infants to correctly interpret the next event.2

Finally, as standard tasks demonstrate, reconciliation between the OF and PR systems is not possible if the two objects are both in the hiding location as the first event comes to an end. Returning to our occlusion event, although the PR system may signal that two objects are present behind the screen, the OF system will not be able to act on this prompt and open a file for the second object because it lacks a unique categorical descriptor or tag for the object. As stated above, the OF system can use only contrastive categorical information to individuate two objects: It cannot use spatiotemporal featural differences (e.g., one object is resting on the left behind the screen, and the other object is resting on the right) or identity featural differences (e.g., one object is a large yellow ball with red dots, and the other object is a small green ball with pink stripes) to establish separate representations for the objects. This explains infants’ catastrophic failure at standard tasks: Because the OF and PR systems disagree about the number of objects present in the first event and this disagreement cannot be reconciled, no coherent representation of the event can be stored in memory, leaving infants with no expectation about what they will see in the next event.

According to the two-system account, the best way to make sense of infants’ successes and failures at standard and modified individuation tasks is thus to understand how the OF and PR systems operate during infancy, what information they each represent, how they exchange information as events unfold, and how these exchanges mature over time.

As is hopefully clear from the preceding sections, the two-system account incorporates elements from both the kind and mapping accounts but differs from them in important ways. In the present research, we focused on the competing claims of the kind and two-system accounts about why contrastive categorical information is critical for success in a standard task. According to the kind account, when the two objects emerge in alternation in the first event, infants must decide whether they are watching a single object with changing properties or two different objects with stable properties. Being able to assign the two objects to different kinds, either spontaneously or via lexical or functional manipulations, enables infants to select the latter option. Members of a kind share intrinsic, stable properties, often with rich conceptual content, and they cannot spontaneously change kinds. Assigning the objects to different kinds thus enables infants to infer that two objects are present in the first event, leading them to expect two objects in the second event.

According to the two-system account, in contrast, infants do not need to establish a kind representation for an object to understand that it cannot spontaneously change its properties: The PR system’s principle of persistence stipulates that all other things being equal, an object will persist as it is, with all of its properties. Thus, during the first event in a standard task, the PR system is able to establish separate representations for the two objects as long as they differ in (a) one or more categorical descriptors or (b) one or more features that have been identified as causally relevant for the event category involved. When the event comes to an end, however, a coherent representation of the event can be stored in memory only if the OF and PR systems agree that two objects are present. Because the OF system relies on categorical information for individuation, access to contrastive descriptors—any types of contrastive descriptors, whether they involve kinds or not—provides the OF system with distinct tags or handles for individuating the two objects, making possible agreement with the PR system about the number of objects present.

The preceding descriptions highlight a clear difference between the kind and two-system accounts. According to the kind account, infants succeed at a standard task only when they can assign the two objects to different kinds with stable, inherent properties. According to the two-system account, however, any type of contrastive categorical information should lead to success, by supplying the OF system with distinct tags for the two objects. In line with this analysis, the present research examined whether infants would succeed at a standard-violation task when the categorical descriptors available to the OF system involved temporary or scant information. We used two different manipulations, one focusing on temporary descriptors (Experiment 1) and one on scant descriptors (Experiment 2).

Experiment 1 built on physical-reasoning findings showing that the PR system assigns categorical, event-specific roles to objects in events (e.g., support, supportee; occluder, occludee; hitter, hittee; for a review, see Lin et al., 2022). To start, 14-month-old infants saw two support events in which two objects from the same basic-level category, two cylinders that differed only in color and pattern, played different roles in relation to a base. In one event, one cylinder was supported by the base, and in the other event, the other cylinder supported the base. Next, the two cylinders emerged in alternation from behind a large screen, which was then lowered to reveal only one of the cylinders. We reasoned that if the OF system (a) retrieved the cylinders’ prior event roles as they emerged from behind the screen (e.g., formerly a support, formerly a supportee), and (b) could take advantage of these descriptors to individuate the cylinders, then infants should succeed at the task. Unlike kind descriptors, role-based descriptors often do not refer to objects’ inherent, stable properties; rather, they are temporary, ephemeral descriptors that change as objects participate in successive events from different categories (e.g., first a supportee, then an occludee). Positive findings would thus support the two-system account by showing that even temporary categorical descriptors can induce success in a standard task.

Experiment 2 built on object-segregation findings showing that young infants who are unable to parse a display composed of two adjacent objects can succeed at doing so if they are first exposed to a static array of three objects from the same basic-level category as one of the adjacent objects (Dueker et al., 2003, Needham and Baillargeon, 2000, Needham et al., 2005). Xu and Carey (2000) argued that static arrays are too limited to result in kind representations, as they provide no lexical, functional, or other information to anchor such representations. Static arrays can result only in “perceptual categories determined by properties” (p. 293) or in “experientially derived shape representations [that] contribute to object individuation at very early stages of perceptual processing” (p. 295). Array-based representations may thus induce success in simple segregation tasks but not in more challenging standard individuation tasks, where kind representations are required. To evaluate these claims, Experiment 2 tested 9-month-old infants with a standard task using two objects from different basic-level categories, which infants this age do not yet spontaneously encode (e.g., a block and a cylinder). Infants first saw two static arrays, one per category, composed of three similarly shaped objects (e.g., three different blocks in one trial, three different cylinders in another trial). Next, the middle objects from the two arrays emerged in alternation from behind a large screen, which was then lowered to reveal only one of the objects. We reasoned that if (a) infants were able to form a categorical representation, however scant or shallow, of each array (e.g., block-shaped objects, cylinder-shaped objects) and (b) the OF system could use these array-based descriptors to individuate the two objects as they emerged from behind the screen, then infants should succeed at the task. Here again, positive findings would support the two-system account by showing that even scant categorical descriptors can induce success in a standard task.

In sum, the present research tested infants with a standard-violation task using two novel types of categorical descriptors for the objects shown in the first event: temporary, role-based descriptors derived from the objects’ causal roles in preceding events, and scant, array-based descriptors derived from the objects’ prior inclusion in static arrays of similarly shaped objects. Across experiments, we tested 14-month-olds using two objects from the same basic-level category, as in Xu et al. (2004), and 9-month-olds using two objects from different basic-level categories, as in Xu and Carey (1996); this made for a richer investigation, with different ages and different standard tasks infants typically fail at these ages. Success in each task would support the two-system account’s claim that the OF system will use any type of contrastive categorical information at its disposal to individuate objects. Such an eclectic approach could only be advantageous for a cognitive system that depends on categorical information to individuate objects: The more types of categorical information it can recruit, the better its performance overall.