Man is the measure of all things. - Protagoras

The assertion by the ancient Greek philosopher Protagoras highlights the notion that reality is defined by how the world is perceived by humans. A contemporary interpretation of this statement is the embodied theory of cognition (e.g. Chemero, 2013; Gallagher, 2017; Yu et al., 2005; Wilson, 2002; Varela et al., 2017), which, diverging from the belief that size and shape are inherent object features (e.g. Op de Beeck et al., 2008; Konkle and Oliva, 2011), posits that human body scale (e.g. size) constrains the perception of objects and the generation of motor responses. For instance, humans evaluate the climbability of steps based on their leg length (Mark, 1987; Warren, 1984), and determine the navigability of apertures according to the critical aperture-to-shoulder-width ratio (Warren and Whang, 1987). Additionally, grasping strategies have been shown to be contingent upon object size relative to one’s body (Cesari and Newell, 2000; Newell et al., 1989) or hand size (Castiello et al., 1993; Tucker and Ellis, 2004). However, the question of how object perception is influenced by the relative size of objects in relation to the human body remains open. Specifically, it is unclear whether this relative size simply acts as a continuous variable for locomotion reference, or if it affects differentiating and organizing object representation based on their ensued affordances.

To underscore the latter point, Gibson, 1979, the pioneer of embodied cognition research, stated that “Detached objects must be comparable in size to the animal under consideration if they are to afford behavior (p.124).” This implies that an object’s affordance, encompassing all action possibilities offered to an animal, is determined by the object’s size relative to the animal’s size rather than its real-world size. For instance, in a naturalistic environment, such as a picnic scene shown in Figure 1a, there may exist a qualitative distinction between objects within (the objects with warm tints in Figure 1a) and beyond (those with cold tints) the size range of humans. Only objects within the range, such as the apple, the umbrella, and the bottles, may afford actions, while those beyond this range, such as the trees and the tent, are largely viewed as part of the environment. Consequently, visual perception may be ecologically constrained, and the body may serve as a metric that facilitates meaningful engagement with the environment by differentiating objects that are accessible for interactions from those not. Further, grounded cognition theory (see Barsalou, 2008 for a review) suggests that the outputs of such differentiation might transcend sensorimotor processes and integrate into supramodal concepts and language. From this perspective, we proposed two hypotheses: first, the affordance of objects will exhibit a qualitative difference between objects within and beyond the size range of an organism’s body; second, affordance-related neural activity will emerge exclusively for objects within the organism’s size range.

Figure 1 with 1 supplement see all

Download asset Open asset

To test these hypotheses, we first measured the affordance of a diverse array of objects varying in real-world sizes (e.g. Konkle and Oliva, 2011). We found a dramatic decline in affordance similarity between objects within and beyond the human body size range, as these objects afforded distinct sets of action possibilities. Notably, the affordance boundary varied in response to the imagined body sizes and showed supramodality. It could also be attained solely through language, as evidenced by the large language model (LLM), ChatGPT (OpenAI, 2023). A subsequent fMRI experiment corroborated the qualitative difference in affordances demarcated by the body size, as affordances of objects within humans’ size range, but not those beyond, were represented in both dorsal and ventral visual streams of the brain. This study advances our understanding of the role of body size in shaping object representation and underscores the significance of body size as a metric for determining object affordances that facilitates meaningful engagement with the environment.

To illustrate how human body size affects object affordances with different sizes, we first characterized the affordances of a set of daily objects. In each trial, we presented a matrix consisting of nine objects and asked participants to report which objects afforded a specific action (e.g. sit-able: a chair, a bed, a skateboard, but not a phone, a laptop, an umbrella, a kettle, a plate, or a hammer; Figure 1b). In this task, there were 14 actions commonly executed in daily life and 24 object images from the THINGS database (Hebart et al., 2019), with sizes ranging from size rank 2 to 8 according to Konkle and Oliva, 2011’s classification. These objects covered real-world sizes from much smaller (17 cm on average, rank 2) to orders of magnitude larger (5317 cm on average, rank 8) than the human body size (see Materials and methods for details). Consequently, affordances for each object were indexed by a 14-dimensional action vector, with the value for each dimension representing the percentage of participants who agreed on a certain action being afforded by the object (e.g. 88% for the action of grasping on a hammer indicating 88% of participants agreed that a hammer affords grasping). Figure 1—figure supplement 1 shows the affordances of two example objects.

An affordance similarity matrix was then constructed where each cell corresponded to the similarity in affordances between a pair of objects (Figure 1c). A clustering analysis revealed a two-cluster structure. Visual inspection suggested that the upper-left cluster consisted of objects smaller than human body size (red labels), and the lower-right cluster contained objects larger than human body size (green labels). Critically, the between-cluster similarity in the affordance similarity matrix approached zero, suggesting a division in affordances located near the body size. To quantify this observation, we calculated the similarity in affordances between each neighboring size rank. Indeed, we identified a clear trough in affordance similarity, dropping to around zero, between size rank 4 (77cm on average) and 5 (146cm on average), which was significantly smaller than that between size rank 3 and 4 (Z=3.91, p<0.001) and that between size rank 5 and 6 (Z=1.66, p=0.048). This trough suggested an affordance boundary between size rank 4 and 5, while affordance similarities between neighboring ranks remained high (rs >0.45) and did not significantly differ from each other (ps >0.05, all ${BF}_{10}$ < 10) on either side of the boundary (Figure 1d, left panel, green lines). This pattern was evident for both genders, indicating no gender difference. Note that the abrupt change in affordance similarity across the boundary cannot be explained by changes in objects’ real-world size, as the similarity in objects’ real-world size was relatively stable across ranks, without any trend of a trough-shape curve (Figure 1d, left panel, yellow line). Intriguingly, rank 4 and rank 5 correspond to 80 cm to 150 cm, a boundary situated between these two ranks is within the range of the body size of a typical human adult. This finding suggested that objects were classified into two categories based on their affordances, with the boundary aligning with human body size.

To better locate the boundary, we focused on the affordance similarity between individual objects within size rank 3–6 (approximately ranging from 30cm to 220 cm in real-world size, the area with grey shade in Figure 1d), where the trough-shape curve was identified. Specifically, we traversed all pairs of objects with similar real-world diagonal sizes (from either the same rank or from neighboring ranks), calculated their average real-world size as an index of the approximate location of the boundary between this pair of objects, and plotted the affordance similarity against the average real-world size of each object pair. As shown in the inset (grey box) of Figure 1d, consistent with the rank-wise analysis, the abrupt decrease in affordance similarity exclusively happened between objects from size rank 4 and 5 (light green dots). The averaged real-world size in these object pairs was 104 cm (95% CI, 105–130 cm) and the affordance similarity in such object pairs was around zero. This result further narrowed the location estimation of the boundary, and demonstrated that the affordance boundary persisted at the level of individual objects.

One may argue that the location of the affordance boundary coincidentally fell within the range of human body size, rather than being directly influenced by it. To rule out this possibility, we directly manipulated participants’ body schema, referring to an experiential and dynamic functioning of the living body within its environment (Merleau-Ponty and Smith, 1962). This experimental approach was able to establish a causal link between body size and affordance boundary, as other potential factors remained constant. Specifically, we instructed a new group of participants to imagine themselves as small as a cat (typical diagonal size: 77 cm, size rank 4, referred to as the ‘cat condition’), and another new group to envision themselves as large as an elephant (typical diagonal size: 577 cm, size rank 7, referred to as the ‘elephant condition’) throughout the task (Figure 2a). A between-subject design was adopted to minimize contamination between conditions. This manipulation was effective, as evidenced by the participants’ reported imagined heights in the cat condition being 42 cm (SD=25.6) and 450 cm (SD=426.8) in the elephant condition on average, respectively, when debriefed at the end of the task.

Figure 2

Download asset Open asset

With exactly the same set of objects, a distinct shift in the affordance boundary was observed for each condition (Figure 2b). In the cat condition, the affordance boundary was identified between size rank 3 and 4, with affordance similarity between size rank 3 and 4 being significantly lower than that between size rank 2 and 3 (Z=1.76, p=0.039) and that between size rank 4 and 5 (Z=1.68, p=0.047). In contrast, in the elephant condition, the affordance boundary shifted to the right, as demonstrated by a decrease in affordance similarity between size rank 6 and 7, and that between size rank 7 and 8 as compared to that between size rank 5 and 6, with a trend towards significance (with size rank 6–7: Z=1.28, p=0.099; with size rank 7–8: Z=1.48, p=0.069). The observation that the affordance boundary shifted to the left under the cat condition and to the right under the elephant condition suggests that affordance perception is influenced even by imagined body size. Furthermore, the cognitive penetrability (Pylyshyn, 1999) of affordance perception implies potential susceptibility of affordance perception to semantic or conceptual transformation or modification.

To test the further speculation that the affordance boundary can be derived solely from conceptual knowledge without direct sensorimotor experience, we employed a disembodied agent, the large language model (LLM) ChatGPT (Chat Generative Pre-trained Transformer; https://openai.com/blog/chatgpt/). This model was trained on a massive corpus of language materials originated from humans, yet it can not receive any sensorimotor information from the environment. Here we asked whether language alone would be sufficient to form an affordance boundary in ChatGPT models as well as in smaller LLMs, BERT (Devlin et al., 2018), and GPT-2 (Radford et al., 2018).

The experimental procedure was similar to that conducted with human participants, except that images were replaced by the corresponding words (see Methods). Given the randomness embedded in response generation, each model was tested 20 times to simulate the sampling of human participants. We found that the affordance similarity curves demonstrated by the ChatGPT models were both trough-shaped between size rank 4 and 5, the same location where the boundary emerged in human participants (Figure 2c, green lines). Further statistical analyses showed a significant difference in affordance similarity between size rank 3 and 4, and that between size rank 4 and 5 (ChatGPT (GPT-3.5): Z=1.98, p=0.024; ChatGPT (GPT-4): Z=2.73, p=0.003). The affordance similarity between size rank 4 and 5 was also lower than that between size rank 5 and 6, yet the difference did not reach the significance (ChatGPT (GPT-3.5): Z=0.96, p=0.17; ChatGPT (GPT-4): Z=1.27, p=0.10). In contrast, no trough-shaped boundary was observed in either BERT or GPT-2 (Figure 2c, yellow lines), despite an apparent but non-significant decrease in affordance similarity in GPT-2 between size rank 5 and 6 (ps >0.20). To further quantify the magnitude of the decrease in affordance similarity between the size rank 4 and 5, we measured the decrease by subtracting the similarity value at the trough from the neighboring similarity values and then subjected it to a permutation test (see Materials and methods). We found a significant decrease in affordance similarity in humans (permutation N=5000, ${\displaystyle p\left(T>T_{obs}\right)}$ = 0.015) and ChatGPT (GPT-4) (permutation N=5000, ${\displaystyle p\left(T>T_{obs}\right)}$ = 0.046), a marginal significant decrease in ChatGPT (GPT-3.5) (permutation N=5000, ${\displaystyle p\left(T>T_{obs}\right)}$ = 0.061), and no significance in either BERT or GPT-2 (ps >0.46, Figure 2d). Thus, the affordance boundary can be derived from language solely without sensorimotor information from environment. Interestingly, it appears to spontaneously emerge when the language processing ability of the LLMs surpasses a certain threshold (i.e. GPT-2 /BERT < ChatGPT models).

A further analysis on the affordances separated by the boundary revealed that objects within human body size range were primarily subjected to hand-related actions such as grasping, holding, and throwing. These affordances typically involve object manipulation with humans’ effectors. In contrast, objects beyond the size range of human body predominantly afforded actions such as sitting and standing, which typically require locomotion or posture change of the whole body around or within the objects. The distinct categories of reported affordances demarcated by the boundary imply that the objects on either side of the boundary may be represented differently in the brain. We thus speculated that the observed behavioral discontinuity is likely underpinned by distinct neural activities, which give rise to these discrete ‘representations’ separated by the boundary.

To test this speculation, we ran an fMRI experiment with a small number of participants to preliminarily investigate the neural basis of the affordance boundary in the brain by measuring neural activity in the dorsal and ventral visual streams when participants were instructed to evaluate whether an action was affordable by an object (Figure 3a). Four objects were chosen from the behavioral experiment: two within the body size range (i.e. bottle and football, WITHIN condition) and the two beyond (i.e. bed and piano, BEYOND condition). Accordingly, four representative actions (to grasp, to kick, to sit, and to lift) were selected in relation to the respective objects. During the scan, the participants were asked to decide whether a probe action was affordable (e.g. grasp-able – bottle, Congruent condition) or not (e.g. sit-able – bottle, Incongruent condition) by each subsequently presented object. The congruency effect, derived from the contrast of Congruent versus Incongruent conditions, is a well-established measure of affordance processing (e.g. Kourtis et al., 2018).

Figure 3 with 3 supplements see all

Download asset Open asset

We examined the congruency effect in two object-selective regions defined by the contrast of objects against baseline (see Materials amd methods), each representing a corresponding visual stream: the posterior fusiform (pFs) in the ventral stream, which is involved in object recognition (e.g. Grill-Spector et al., 2000; Malach et al., 1995) and objects’ real-world size processing (Konkle and Oliva, 2012; Snow et al., 2011), and the superior parietal lobule (SPL) in the dorsal stream, one of the core tool network regions (e.g. Filimon et al., 2007; Matić et al., 2020). For the rest object-selective regions identified in this experiment, see Figure 3—figure supplement 1 and Supplementary file 1a. A repeated-measures ANOVA with object type (WITHIN versus BEYOND) and congruency (Congruent versus Incongruent) as within-subject factors was performed for each ROI, respectively. A significant interaction between object type and congruency was observed in both ROIs (SPL: F(1,11) = 15.47, p=0.002, ${\eta }^2$ = .58; pFs: F(1,11) = 24.93, p<0.001, ${\eta }^2$ = .69), suggesting that these regions represented affordances differentially based on object type (Figure 3c). A post hoc simple effect analysis revealed the congruency effect solely for objects within body size range (SPL: p<0.001; pFs: p=0.021), not for objects beyond (ps >0.41). In addition, the main effect of object type was not significant in either ROI (ps >0.17), suggesting that the absence of the congruency effect for objects beyond the body size cannot be attributable to compromised engagement in viewing these objects. In addition, a whole-brain analysis was performed, and no region showed the congruency effect for the objects beyond the body size. Taken together, the affordance boundary not only separated the objects into two categories based on their relative size to human body, but also delineated the range of objects that evoked neural representations associated with affordance processing.

In addition to the pFs and SPL, we also examined the congruency effect in the lateral occipital cortex (LO), which is involved in object representation (e.g. Grill-Spector et al., 2000; Konkle and Caramazza, 2013) and provides inputs to both the pFs and SPL (Hebart et al., 2018). Meanwhile, the primary motor cortex (M1), which receives inputs from the dorsal stream (Vainio and Ellis, 2020), is involved in affordance processing (e.g. McDannald et al., 2018) and action executions (Binkofski et al., 2002). Although both the LO and M1 showed a significantly higher response to objects than baseline, no congruency effect in affordance for objects within the body size was observed (main effect of congruency: F(1,11) = 1.74, p=0.214, ${\eta }^2$ = 0.13, Figure 3—figure supplement 2). Therefore, it is unlikely that the representation of affordance is exclusively dictated by visual inputs or automatically engaged in motor execution. This finding suggests that affordance perception likely requires perceptual processing and is not necessarily reflected in motor execution, diverging from Gibsonian concept of direct perception.

One long-lasting debate on affordance centers on the distinction between representational and direct perception of affordance. An outstanding theme shared by many embodied theories of cognition is the replacement hypothesis (e.g. van Gelder, 1998), which challenges the necessity of representation as posited by computationalism’s cognitive theories (e.g. Fodor, 1975). This hypothesis suggests that input is discretized/categorized and subjected to abstraction or symbolization, creating discrete stand-ins for the input (e.g. representations/states). Such representationalization would lead to a categorization between the affordable (the objects) and those beyond affordance (the environment), in contrast to the perspective offered by embodied theories. The present study probed this ‘representationalization’ of affordance by examining whether affordance perception introduces discontinuity and qualitative dissociation in response to continuous action-related physical features (such as object size relative to the agents), which allows sensorimotor input to be assigned into discrete states/kinds, in line with the representation-based view under the constraints of body size. Alternatively, it assessed whether activity directly mirrors the input, free from discretization/categorization/abstraction, in line with the representation-free view.

First, our study found evidence demonstrating discretization in affordance perception. Then, through the body imagination experiment, we provided causal evidence suggesting that this discretization originates from sensorimotor interactions with objects rather than amodal sources, such as abstract object concepts independent of agent motor capability. Finally, we demonstrated the supramodality of this embodied discontinuity by leveraging the recent advances in AI. We showed that the discretization in affordance perception is supramodally accessible to disembodied agents such as large language models (LLMs), which lack sensorimotor input but can access linguistic materials built upon discretized representations. These results collectively suggest that sensorimotor input undergoes discretization, as implied in the computationalism’s idea of representation. Note that, these results are not contradictory to the claim of the embodied theories, as these representations do shape processes beyond the sensorimotor domain but after discretization.

This observed boundary in affordance perception extends the understanding of the discontinuity in perception in response to the continuity of physical inputs (Harnad, 1987; Young et al., 1997). Perceptual boundary has been demonstrated in various domains, such as color perception (Bornstein and Korda, 1984), speech-sounds perception (Liberman et al., 1957), and facial gender discrimination (Campanella et al., 2001). The boundaries reflect a fundamental adaptation of perception to facilitate categorizations necessary for an organism (Goldstone and Hendrickson, 2010). Our study, for the first time, unveiled a boundary in object affordance, wherein affordance similarity across the boundary was significantly lower than that within the boundary. Critically, the boundary separating object affordances along a size axis coincided with human body size, suggesting that object affordances are characterized in a dimension scaled by human body size.

What is the function of the affordance boundary? About four decades ago, Gibson, 1979 postulated that only objects of sizes comparable to an animal’s body size are amenable to interaction and capable of providing affordances to the animal, thereby possessing ecological values that distinctly differ from those of larger objects. In this study, we expand upon this notion by arguing that the affordance boundary serves to delineate (manipulable) objects from their surrounding environment. In other words, objects within the range of an animal’s body size are indeed objects in the animal’s eye and possess affordances as defined by Gibson. In contrast, objects larger than that range typically surpass the animal’s motor capabilities, rendering them too cumbersome for effective manipulation. Consequently, these larger objects are less likely to be considered as typical targets for manipulation by the animal, as opposed to the smaller objects. That is, they are perceived not as the “objects” in the animal’s eye, but as part of the background environment, due to their impracticality for direct interactions. Future studies should incorporate a broader range of objects and a more comprehensive set of affordances for finer delineation of the representational discontinuity between objects and the environment.

This speculation aligns with previous fMRI studies where large objects activated the medial portion of the ventral temporal cortex (Huang et al., 2022; Magri et al., 2021), overlapping with the parahippocampus gyrus involved in scene representation (Park et al., 2011; Troiani et al., 2014), and smaller objects activated the lateral portion, such as the pFs, where the congruency effect of affordance was identified in our study. Furthermore, we found that the congruency effect was only evident for objects within the body size range, but not for objects beyond, supporting the idea that affordance is typically represented only for objects within the body size range. While it is acknowledged that the sample size of the fMRI study was small (12 participants), necessitating cautious interpretation of its results, the observed neural-level affordance discontinuity is notable. That is, qualitative differences in neural activity between objects within the affordance boundary and those beyond replicated our behavioral findings. This convergent evidence reinforced our claim that objects were discretized into two broad categories along the continuous size axis, with affordance only being manifested for objects within the boundary.

In this context, an animal’s body size and sensorimotor capacity determine the boundary of manipulation, and thus, the boundary between manipulable objects and the environment. Therefore, our study provides a novel perspective on a long-standing question in psychology, cognitive science, and philosophy: what constitutes an object? Existing psychological studies, especially in the field of vision, define objects in a disembodied manner, primarily relying on their physical properties such as shape (e.g. Op de Beeck et al., 2008) and absolute size (e.g. Konkle and Oliva, 2011). Our identification of the affordance boundary presents a new source of object-ness: the capability of being a source of affordance under the constraints of an animal’s sensorimotor capacity, which resonates the embodied influence on the formation of abstract concepts (e.g. Barsalou, 1999; Lakoff and Johnson, 1980) of objects and environment. Consistently, our fMRI data did not show the congruency effect for objects beyond the body size range, distinct from objects within this range, suggesting a categorization influenced by objects’ relative size to the human body. In this respect, man is indeed the measure of all things.

The metric provided by the body size, however, was changeable when the body schema was intentionally altered through participants’ imagination of possessing either a cat- or elephant-sized body, with which the participants had no prior sensorimotor experience. Importantly, they perceived new affordances in a manner as if they had had embodied experience with this new body schema. Therefore, this finding suggests that the affordance boundary is cognitively penetrable, arguing against the directness of affordance perception (e.g. Gibson, 1979; Greeno, 1994; Prindle et al., 1980) or the exclusive sensorimotor origin of affordances (e.g. Gallagher, 2017; Thompson, 2010; Hutto and Myin, 2012; Chemero, 2013). Further, this finding that the boundary adapted to manipulation on body schema suggests that the abstraction/representationalization may be dynamically updated in response to the current motor capacity and body schema of the agent, suggesting that the affordance-based process is probably determined dynamically by the nature of the agent-object dyads, rather than being a fixed belief about objects. Future studies could explore the dynamics of affordance representationalization, probably by investigating how affordance representations evolve during active interactions with novel objects or under conditions of altered motor capabilities. Finally, our findings also suggest that disembodied conceptual knowledge pertinent to action likely modulates affordance perception. Indeed, it has been proposed that conceptual knowledge is grounded in the same neural system that supports action (Barsalou, 1999; Glenberg et al., 2013; Wilson and Golonka, 2013), thereby suggesting that sensorimotor information, along with other model inputs, may be embedded in language (e.g. Casasanto, 2011; Glenberg and Gallese, 2012; Stanfield and Zwaan, 2001), as the grounded theory proposed (see Barsalou, 2008 for a review).

Direct evidence for this speculation comes from the disembodied ChatGPT models, which showed an evident affordance boundary despite lacking direct interaction with the environment. We speculated that ChatGPT models may have formed the affordance boundary through a human prism ingrained within its linguistic training corpus. In fact, when inquired about the size of a hypothetical body constructed for its use, ChatGPT (GPT-4) replied, “It could be the size of an average adult human, around 5 feet 6 inches (167.6 cm) tall. This would allow me to interact with the world and people in a familiar way.” Critically, this size corresponds to the location where the affordance boundary of ChatGPT models was found. In essence, a virtual body schema may have automatically emerged in ChatGPT models, possibly based on the body schema inherited from humans through language, enabling ChatGPT models to display a preliminary ability to reason the relationship between bodily action and objects. It should be noted that the affordance boundary was not present in all LLMs tested. Specifically, LLMs with a smaller number of parameters, such as BERT and GPT-2, did not exhibit any robust boundary, suggesting the emergence of the boundary may depend on language processing ability determined by the scale of training datasets and the complexity of the model (Hestness et al., 2017; Brown et al., 2020), as well as alignment methods used in fine-tuning the model (Ouyang et al., 2022). Nevertheless, caution should be taken when interpreting the capability of LLMs like ChatGPT, which are often considered ‘black boxes’. That is, our observation indicates that certain sensorimotor information is embedded within human language materials presumably through linguistic statistics, but it is not sufficient to assert that LLMs have developed a human-like ability to represent affordances. Furthermore, such information alone may be insufficient for LLMs to mimic the characteristics of the affordance perception in biological intelligence. Future studies are needed to elucidate such limitations.

While the primary focus of our study concerns the nature of human perception of affordance, our findings on ChatGPT models raise an intriguing question that extends beyond psychology and neuroscience into the domain of artificial intelligence (AI). The AI field has predominantly concentrated on disembodied cognition, such as vision and language. In contrast, the utilization of sensorimotor information to interact with and adapt to the world, including affordance perception in our study, represents a crucial human cognitive achievement that remains elusive for AI systems. Traditional AI (i.e. task-specific AI) has been confined to narrowly defined tasks, with substantial limitations in adaptability and autonomy. Accordingly, these systems have served primarily as tools for humans to achieve specific outcomes, rather than as autonomous agents capable of independently formulating goals and translating them into actionable plans. In recent years, significant efforts have been directed towards evolving traditional AI into more agent-like entities, especially in domains like navigation, object manipulation, and other interactions with the physical world. Despite these advancements, the capabilities of AI still fall behind human-level intelligence. On the other hand, embodied cognition theories suggest that sensorimotor interactions with the environment are foundational for various cognitive domains. From this point of view, endowing AI with human-level abilities in physical agent-environment interactions might provide an unreplaceable missing piece for achieving Artificial General Intelligence (AGI). This development would significantly facilitate AI’s role in robotics, particularly in actions essential for survival and goal accomplishment, a promising direction for the next breakthrough in AI (Gupta et al., 2021; Smith and Gasser, 2005).

However, equipping a disembodied AI with the ability for embodied interaction planning within a specific environment remains a complex challenge. By testing the potential representationalization of action possibilities (affordances) in both humans and LLMs, the present study suggests a new approach to enhancing AI’s interaction ability with the environment. For instance, our finding of supramodal affordance representation may indicate a possible pathway for disembodied LLMs to engage in embodied physical interactions with their surroundings. From an optimistic view, these results suggest that LLM-based agents, if appropriately designed, may leverage affordance representations embedded in language to interact with the physical world. Indeed, by clarifying and aligning such representations with the physical constitutes of LLM-based agents, and even by explicitly constructing an agent-specific object space, we may foster the sensorimotor interaction abilities of LLM-based agents. This progression could lead to achieving animal-level interaction abilities with the world, potentially sparking new developments in the field of embodied cognition theories.

Although our study showed the supramodality of the representationalization of affordance, two questions remain. First, the magnitude of the boundary observed in ChatGPT models was smaller than that in humans. This discrepancy might be compensated by merely enhancing the language processing ability of LLMs. Alternatively, direct interaction with the environment may be necessary for LLMs to achieve human-level performance in affordance perception. Second, the size of virtual body schema of ChatGPT models, if present, coincided with human body size. Integrating LLMs with real robots (e.g. Driess et al., 2023) may pose a challenge because the to-be-supported robots or cars for autopilot might not fall within human body size range. Future studies may be needed to align the inherited body schema with the actual constitution of the robots. Addressing these questions is beyond the scope of the present study but may hold significant implications for the development of AI systems possessing human-level ingenuity and adaptability in interacting with the world.

In summary, our findings regarding the affordance boundary highlight the interdependence between an agent and the external world in shaping cognition. Furthermore, taking our finding with embodied humans and disembodied LLMs into account, we propose a revision to the purely sensorimotor-based concept of affordance by emphasizing a disembodied, perhaps conceptual, addition to it. That is, the embodied cognition and symbolic processing of language may be more intricately and fundamentally connected than previously thought: perception-action problems and language problems can be treated as the same kind of process (Wilson and Golonka, 2013). In this context, man is the measure of both the world and the words, for both humans and AIs. The presence of such a metric may shed light on the development of AI systems that can fully capture essential human abilities founded on sensorimotor interactions with the world.

All analyses are included in the manuscript. The data are freely available from Figshare.

1. 1. Feng XR
   2. Xu S
   3. Li Y
   4. Liu J

   (2024) figshare

   Data for 'Body size as a metric for the affordable world'.

   https://doi.org/10.6084/m9.figshare.25248985.v1

1. 1. Barsalou LW

   (1999) Perceptual symbol systems

   The Behavioral and Brain Sciences 22:577–609.

   https://doi.org/10.1017/s0140525x99002149

   • PubMed
   • Google Scholar
2. 1. Barsalou LW

   (2008) Grounded cognition

   Annual Review of Psychology 59:617–645.

   https://doi.org/10.1146/annurev.psych.59.103006.093639

   • PubMed
   • Google Scholar
3. 1. Beckmann CF
   2. Jenkinson M
   3. Smith SM

   (2003) General multilevel linear modeling for group analysis in FMRI

   NeuroImage 20:1052–1063.

   https://doi.org/10.1016/S1053-8119(03)00435-X

   • PubMed
   • Google Scholar
4. 1. Binkofski F
   2. Fink GR
   3. Geyer S
   4. Buccino G
   5. Gruber O
   6. Shah NJ
   7. Taylor JG
   8. Seitz RJ
   9. Zilles K
   10. Freund HJ

   (2002) Neural activity in human primary motor cortex areas 4a and 4p is modulated differentially by attention to action

   Journal of Neurophysiology 88:514–519.

   https://doi.org/10.1152/jn.2002.88.1.514

   • PubMed
   • Google Scholar
5. 1. Borghi AM

   (2005)

   Grounding Cognition: The Role of Perception and Action in Memory, Language, and Thinking

   Object concepts and action, Grounding Cognition: The Role of Perception and Action in Memory, Language, and Thinking, Cambridge University Press.

   • Google Scholar
6. 1. Bornstein MH
   2. Korda NO

   (1984) Discrimination and matching within and between hues measured by reaction times: some implications for categorical perception and levels of information processing

   Psychological Research 46:207–222.

   https://doi.org/10.1007/BF00308884

   • PubMed
   • Google Scholar
7. Preprint

   1. Brown TB
   2. Mann B
   3. Ryder N
   4. Subbiah M
   5. Kaplan J
   6. Dhariwal P
   7. Amodei D

   (2020) Language Models Are Few-Shot Learners

   arXiv.

   https://arxiv.org/abs/2005.14165

   • Google Scholar
8. 1. Campanella S
   2. Chrysochoos A
   3. Bruyer R

   (2001) Categorical perception of facial gender information: Behavioural evidence and the face-space metaphor

   Visual Cognition 8:237–262.

   https://doi.org/10.1080/13506280042000072

   • Google Scholar
9. 1. Casasanto D

   (2011) Different bodies, different minds: the body specificity of language and thought

   Current Directions in Psychological Science 20:378–383.

   https://doi.org/10.1177/0963721411422058

   • Google Scholar
10. 1. Castiello U
    2. Bennett KMB
    3. Stelmach GE

    (1993) Reach to grasp: the natural response to perturbation of object size

    Experimental Brain Research 94:163–178.

    https://doi.org/10.1007/BF00230479

    • PubMed
    • Google Scholar
11. 1. Cesari P
    2. Newell KM

    (2000) Body-scaled transitions in human grip configurations

    Journal of Experimental Psychology. Human Perception and Performance 26:1657–1668.

    https://doi.org/10.1037//0096-1523.26.5.1657

    • PubMed
    • Google Scholar
12. 1. Chemero A

    (2013) Radical embodied cognitive science

    Review of General Psychology 17:145–150.

    https://doi.org/10.1037/a0032923

    • Google Scholar
13. Software

    1. Colling LJ

    (2021) Ljcolling/go-Bayesfactor, version Version v0.9.0

    Zenodo.

    https://zenodo.org/records/4642331
14. Preprint

    1. Devlin J
    2. Chang MW
    3. Lee K
    4. Toutanova K

    (2018) Bert: Pre-Training of Deep Bidirectional Transformers for Language Understanding

    arXiv.

    https://arxiv.org/abs/1810.04805

    • Google Scholar
15. 1. Diedenhofen B
    2. Musch J

    (2015) cocor: A comprehensive solution for the statistical comparison of correlations

    PLOS ONE 10:e0121945.

    https://doi.org/10.1371/journal.pone.0121945

    • PubMed
    • Google Scholar
16. Preprint

    1. Driess D
    2. Xia F
    3. Sajjadi MS
    4. Lynch C
    5. Chowdhery A
    6. Ichter B
    7. Florence P

    (2023) PaLM-E: An Embodied Multimodal Language Model

    arXiv.

    https://arxiv.org/abs/2303.03378

    • Google Scholar
17. 1. Fan L
    2. Li H
    3. Zhuo J
    4. Zhang Y
    5. Wang J
    6. Chen L
    7. Yang Z
    8. Chu C
    9. Xie S
    10. Laird AR
    11. Fox PT
    12. Eickhoff SB
    13. Yu C
    14. Jiang T

    (2016) The human brainnetome atlas: a new brain atlas based on connectional architecture

    Cerebral Cortex 26:3508–3526.

    https://doi.org/10.1093/cercor/bhw157

    • Google Scholar
18. 1. Filimon F
    2. Nelson JD
    3. Hagler DJ
    4. Sereno MI

    (2007) Human cortical representations for reaching: mirror neurons for execution, observation, and imagery

    NeuroImage 37:1315–1328.

    https://doi.org/10.1016/j.neuroimage.2007.06.008

    • PubMed
    • Google Scholar
19. Book

    1. Fodor JA

    (1975)

    The Language of Thought

    Harvard University Press.

    • Google Scholar
20. Book

    1. Gallagher S

    (2017)

    Enactivist Interventions: Rethinking the Mind

    Oxford University Press.

    • Google Scholar
21. Book

    1. Gibson JJ

    (1979)

    The Ecological Approach to Visual Perception

    Psychology Press.

    • Google Scholar
22. 1. Glenberg AM
    2. Gallese V

    (2012) Action-based language: A theory of language acquisition, comprehension, and production

    Cortex; a Journal Devoted to the Study of the Nervous System and Behavior 48:905–922.

    https://doi.org/10.1016/j.cortex.2011.04.010

    • PubMed
    • Google Scholar
23. 1. Glenberg AM
    2. Witt JK
    3. Metcalfe J

    (2013) From the revolution to embodiment: 25 years of cognitive psychology

    Perspectives on Psychological Science 8:573–585.

    https://doi.org/10.1177/1745691613498098

    • PubMed
    • Google Scholar
24. 1. Goldstone RL
    2. Hendrickson AT

    (2010) Categorical perception

    Wiley Interdisciplinary Reviews. Cognitive Science 1:69–78.

    https://doi.org/10.1002/wcs.26

    • PubMed
    • Google Scholar
25. 1. Greeno JG

    (1994) Gibson’s affordances

    Psychological Review 101:336–342.

    https://doi.org/10.1037/0033-295x.101.2.336

    • PubMed
    • Google Scholar
26. 1. Grill-Spector K
    2. Kushnir T
    3. Hendler T
    4. Malach R

    (2000) The dynamics of object-selective activation correlate with recognition performance in humans

    Nature Neuroscience 3:837–843.

    https://doi.org/10.1038/77754

    • PubMed
    • Google Scholar
27. 1. Gupta A
    2. Savarese S
    3. Ganguli S
    4. Fei-Fei L

    (2021) Embodied intelligence via learning and evolution

    Nature Communications 12:5721.

    https://doi.org/10.1038/s41467-021-25874-z

    • PubMed
    • Google Scholar
28. Book

    1. Harnad S

    (1987)

    Psychophysical and Cognitive Aspects of Categorical Perception: A Critical Overview

    University of Southampton.

    • Google Scholar
29. 1. Hebart MN
    2. Bankson BB
    3. Harel A
    4. Baker CI
    5. Cichy RM

    (2018) The representational dynamics of task and object processing in humans

    eLife 7:e32816.

    https://doi.org/10.7554/eLife.32816

    • PubMed
    • Google Scholar
30. 1. Hebart MN
    2. Dickter AH
    3. Kidder A
    4. Kwok WY
    5. Corriveau A
    6. Van Wicklin C
    7. Baker CI

    (2019) THINGS: A database of 1,854 object concepts and more than 26,000 naturalistic object images

    PLOS ONE 14:e0223792.

    https://doi.org/10.1371/journal.pone.0223792

    • PubMed
    • Google Scholar
31. Preprint

    1. Hestness J
    2. Narang S
    3. Ardalani N
    4. Diamos G
    5. Jun H
    6. Kianinejad H
    7. Zhou Y

    (2017) Deep Learning Scaling Is Predictable, Empirically

    arXiv.

    https://arxiv.org/abs/1712.00409

    • Google Scholar
32. 1. Huang T
    2. Song Y
    3. Liu J

    (2022) Real-world size of objects serves as an axis of object space

    Communications Biology 5:749.

    https://doi.org/10.1038/s42003-022-03711-3

    • PubMed
    • Google Scholar
33. Book

    1. Hutto DD
    2. Myin E

    (2012) Radicalizing Enactivism: Basic Minds without Content

    MIT Press.

    https://doi.org/10.7551/mitpress/9780262018548.001.0001

    • Google Scholar
34. 1. Jenkinson M
    2. Bannister P
    3. Brady M
    4. Smith S

    (2002) Improved optimization for the robust and accurate linear registration and motion correction of brain images

    NeuroImage 17:825–841.

    https://doi.org/10.1016/s1053-8119(02)91132-8

    • PubMed
    • Google Scholar
35. 1. Jenkinson Mark
    2. Beckmann CF
    3. Behrens TEJ
    4. Woolrich MW
    5. Smith SM

    (2012) FSL

    NeuroImage 62:782–790.

    https://doi.org/10.1016/j.neuroimage.2011.09.015

    • PubMed
    • Google Scholar
36. Preprint

    1. Kay W
    2. Carreira J
    3. Simonyan K
    4. Zhang B
    5. Hillier C
    6. Vijayanarasimhan S
    7. Zisserman A

    (2017) The Kinetics Human Action Video Dataset

    arXiv.

    https://arxiv.org/abs/1705.06950

    • Google Scholar
37. 1. Konkle T
    2. Oliva A

    (2011) Canonical visual size for real-world objects

    Journal of Experimental Psychology. Human Perception and Performance 37:23–37.

    https://doi.org/10.1037/a0020413

    • PubMed
    • Google Scholar
38. 1. Konkle T
    2. Oliva A

    (2012) A real-world size organization of object responses in occipitotemporal cortex

    Neuron 74:1114–1124.

    https://doi.org/10.1016/j.neuron.2012.04.036

    • PubMed
    • Google Scholar
39. 1. Konkle T
    2. Caramazza A

    (2013) Tripartite organization of the ventral stream by animacy and object size

    The Journal of Neuroscience 33:10235–10242.

    https://doi.org/10.1523/JNEUROSCI.0983-13.2013

    • PubMed
    • Google Scholar
40. 1. Kourtis D
    2. Vandemaele P
    3. Vingerhoets G

    (2018) Concurrent cortical representations of function- and size-related object affordances: an fmri study

    Cognitive, Affective, & Behavioral Neuroscience 18:1221–1232.

    https://doi.org/10.3758/s13415-018-0633-1

    • Google Scholar
41. 1. Lakoff G
    2. Johnson M

    (1980) The metaphorical structure of the human conceptual system

    Cognitive Science 4:195–208.

    https://doi.org/10.1207/s15516709cog0402_4

    • Google Scholar
42. 1. Liberman AM
    2. Harris KS
    3. Hoffman HS
    4. Griffith BC

    (1957) The discrimination of speech sounds within and across phoneme boundaries

    Journal of Experimental Psychology 54:358–368.

    https://doi.org/10.1037/h0044417

    • PubMed
    • Google Scholar
43. 1. Magri C
    2. Konkle T
    3. Caramazza A

    (2021) The contribution of object size, manipulability, and stability on neural responses to inanimate objects

    NeuroImage 237:118098.

    https://doi.org/10.1016/j.neuroimage.2021.118098

    • PubMed
    • Google Scholar
44. 1. Malach R
    2. Reppas JB
    3. Benson RR
    4. Kwong KK
    5. Jiang H
    6. Kennedy WA
    7. Ledden PJ
    8. Brady TJ
    9. Rosen BR
    10. Tootell RB

    (1995) Object-related activity revealed by functional magnetic resonance imaging in human occipital cortex

    PNAS 92:8135–8139.

    https://doi.org/10.1073/pnas.92.18.8135

    • Google Scholar
45. 1. Mark LS

    (1987) Eyeheight-scaled information about affordances: a study of sitting and stair climbing

    Journal of Experimental Psychology. Human Perception and Performance 13:361–370.

    https://doi.org/10.1037//0096-1523.13.3.361

    • PubMed
    • Google Scholar
46. 1. Matić K
    2. Op de Beeck H
    3. Bracci S

    (2020) It’s not all about looks: The role of object shape in parietal representations of manual tools

    Cortex; a Journal Devoted to the Study of the Nervous System and Behavior 133:358–370.

    https://doi.org/10.1016/j.cortex.2020.09.016

    • PubMed
    • Google Scholar
47. 1. McDannald DW
    2. Mansour M
    3. Rydalch G
    4. Bolton DAE

    (2018) Motor affordance for grasping a safety handle

    Neuroscience Letters 683:131–137.

    https://doi.org/10.1016/j.neulet.2018.05.040

    • PubMed
    • Google Scholar
48. Book

    1. Merleau-Ponty M
    2. Smith C

    (1962) Phenomenology of Perception

    London: Routledge.

    https://doi.org/10.4324/9780203720714

    • Google Scholar
49. 1. NCD Risk Factor (NCD-RisC)

    (2016) A century of trends in adult human height

    eLife 5:e13410.

    https://doi.org/10.7554/eLife.13410

    • Google Scholar
50. 1. Newell KM
    2. Scully DM
    3. McDonald PV
    4. Baillargeon R

    (1989) Task constraints and infant grip configurations

    Developmental Psychobiology 22:817–831.

    https://doi.org/10.1002/dev.420220806

    • PubMed
    • Google Scholar
51. 1. Op de Beeck HP
    2. Torfs K
    3. Wagemans J

    (2008) Perceived shape similarity among unfamiliar objects and the organization of the human object vision pathway

    The Journal of Neuroscience 28:10111–10123.

    https://doi.org/10.1523/JNEUROSCI.2511-08.2008

    • Google Scholar
52. Website

    1. OpenAI

    (2023) Introducing ChatGPT

    Accessed February 12, 2023.

    https://openai.com/blog/chatgpt
53. 1. Ouyang L
    2. Wu J
    3. Jiang X
    4. Almeida D
    5. Wainwright C
    6. Mishkin P
    7. Lowe R

    (2022)

    Training language models to follow instructions with human feedback

    Advances in Neural Information Processing Systems 35:27730–27744.

    • Google Scholar
54. 1. Park S
    2. Brady TF
    3. Greene MR
    4. Oliva A

    (2011) Disentangling scene content from spatial boundary: complementary roles for the parahippocampal place area and lateral occipital complex in representing real-world scenes

    The Journal of Neuroscience 31:1333–1340.

    https://doi.org/10.1523/JNEUROSCI.3885-10.2011

    • PubMed
    • Google Scholar
55. 1. Pearson K
    2. Filon LNG

    (1898) VII. Mathematical contributions to the theory of evolution.— IV. On the probable errors of frequency constants and on the influence of random selection on variation and correlation

    Philosophical Transactions of the Royal Society of London. Series A, Containing Papers of A Mathematical or Physical Character 191:229–311.

    https://doi.org/10.1098/rsta.1898.0007

    • Google Scholar
56. 1. Prindle SS
    2. Carello C
    3. Turvey MT

    (1980) Animal-environment mutuality and direct perception

    Behavioral and Brain Sciences 3:395–397.

    https://doi.org/10.1017/S0140525X0000563X

    • Google Scholar
57. 1. Pylyshyn Z

    (1999) Is vision continuous with cognition? The case for cognitive impenetrability of visual perception

    The Behavioral and Brain Sciences 22:341–365.

    https://doi.org/10.1017/s0140525x99002022

    • PubMed
    • Google Scholar
58. Book

    1. Radford A
    2. Narasimhan K
    3. Salimans T
    4. Sutskever I

    (2018)

    Improving Language Understanding by Generative Pre-Training

    OpenAI.

    • Google Scholar
59. 1. Sakreida K
    2. Effnert I
    3. Thill S
    4. Menz MM
    5. Jirak D
    6. Eickhoff CR
    7. Ziemke T
    8. Eickhoff SB
    9. Borghi AM
    10. Binkofski F

    (2016) Affordance processing in segregated parieto-frontal dorsal stream sub-pathways

    Neuroscience & Biobehavioral Reviews 69:89–112.

    https://doi.org/10.1016/j.neubiorev.2016.07.032

    • Google Scholar
60. 1. Smith SM

    (2002) Fast robust automated brain extraction

    Human Brain Mapping 17:143–155.

    https://doi.org/10.1002/hbm.10062

    • PubMed
    • Google Scholar
61. 1. Smith L
    2. Gasser M

    (2005) The development of embodied cognition: six lessons from babies

    Artificial Life 11:13–29.

    https://doi.org/10.1162/1064546053278973

    • PubMed
    • Google Scholar
62. 1. Snow JC
    2. Pettypiece CE
    3. McAdam TD
    4. McLean AD
    5. Stroman PW
    6. Goodale MA
    7. Culham JC

    (2011) Bringing the real world into the fMRI scanner: repetition effects for pictures versus real objects

    Scientific Reports 1:130.

    https://doi.org/10.1038/srep00130

    • PubMed
    • Google Scholar
63. 1. Stanfield RA
    2. Zwaan RA

    (2001) The effect of implied orientation derived from verbal context on picture recognition

    Psychological Science 12:153–156.

    https://doi.org/10.1111/1467-9280.00326

    • PubMed
    • Google Scholar
64. Book

    1. Thompson E

    (2010)

    Mind in Life: Biology, Phenomenology, and the Sciences of Mind

    Harvard University Press.

    • Google Scholar
65. 1. Troiani V
    2. Stigliani A
    3. Smith ME
    4. Epstein RA

    (2014) Multiple object properties drive scene-selective regions

    Cerebral Cortex 24:883–897.

    https://doi.org/10.1093/cercor/bhs364

    • PubMed
    • Google Scholar
66. 1. Tucker M
    2. Ellis R

    (2004) Action priming by briefly presented objects

    Acta Psychologica 116:185–203.

    https://doi.org/10.1016/j.actpsy.2004.01.004

    • PubMed
    • Google Scholar
67. Book

    1. Unpingco J

    (2016) Python for Probability, Statistics, and Machine Learning

    Cham: Springer International Publishing.

    https://doi.org/10.1007/978-3-319-30717-6

    • Google Scholar
68. 1. Vainio L
    2. Ellis R

    (2020) Action inhibition and affordances associated with a non-target object: An integrative review

    Neuroscience & Biobehavioral Reviews 112:487–502.

    https://doi.org/10.1016/j.neubiorev.2020.02.029

    • Google Scholar
69. 1. van Gelder T

    (1998) The dynamical hypothesis in cognitive science

    The Behavioral and Brain Sciences 21:615–628.

    https://doi.org/10.1017/s0140525x98001733

    • PubMed
    • Google Scholar
70. Book

    1. Varela FJ
    2. Thompson E
    3. Rosch E

    (2017) The Embodied Mind

    MIT press.

    https://doi.org/10.7551/mitpress/9780262529365.001.0001

    • Google Scholar
71. 1. Wagenmakers E-J
    2. Wetzels R
    3. Borsboom D
    4. van der Maas HLJ

    (2011) Why psychologists must change the way they analyze their data: the case of psi: comment on Bem (2011)

    Journal of Personality and Social Psychology 100:426–432.

    https://doi.org/10.1037/a0022790

    • PubMed
    • Google Scholar
72. 1. Warren WH

    (1984) Perceiving affordances: visual guidance of stair climbing

    Journal of Experimental Psychology. Human Perception and Performance 10:683–703.

    https://doi.org/10.1037//0096-1523.10.5.683

    • PubMed
    • Google Scholar
73. 1. Warren WH Jr
    2. Whang S

    (1987) Visual guidance of walking through apertures: body-scaled information for affordances

    Journal of Experimental Psychology. Human Perception and Performance 13:371–383.

    https://doi.org/10.1037//0096-1523.13.3.371

    • PubMed
    • Google Scholar
74. 1. Waskom ML

    (2021) seaborn: statistical data visualization

    Journal of Open Source Software 6:3021.

    https://doi.org/10.21105/joss.03021

    • Google Scholar
75. 1. Wilson M

    (2002) Six views of embodied cognition

    Psychonomic Bulletin & Review 9:625–636.

    https://doi.org/10.3758/bf03196322

    • PubMed
    • Google Scholar
76. 1. Wilson AD
    2. Golonka S

    (2013) Embodied cognition is not what you think it is

    Frontiers in Psychology 4:58.

    https://doi.org/10.3389/fpsyg.2013.00058

    • PubMed
    • Google Scholar
77. 1. Worsley KJ

    (2001) Statistical analysis of activation images

    Functional MRI: An Introduction to Methods 14:251–270.

    https://doi.org/10.1093/acprof:oso/9780192630711.001.0001

    • Google Scholar
78. 1. Young AW
    2. Rowland D
    3. Calder AJ
    4. Etcoff NL
    5. Seth A
    6. Perrett DI

    (1997) Facial expression megamix: tests of dimensional and category accounts of emotion recognition

    Cognition 63:271–313.

    https://doi.org/10.1016/s0010-0277(97)00003-6

    • PubMed
    • Google Scholar
79. 1. Yu C
    2. Ballard DH
    3. Aslin RN

    (2005) The role of embodied intention in early lexical acquisition

    Cognitive Science 29:961–1005.

    https://doi.org/10.1207/s15516709cog0000_40

    • PubMed
    • Google Scholar
80. 1. Zhen Z
    2. Yang Z
    3. Huang L
    4. Kong X-Z
    5. Wang X
    6. Dang X
    7. Huang Y
    8. Song Y
    9. Liu J

    (2015) Quantifying interindividual variability and asymmetry of face-selective regions: a probabilistic functional atlas

    NeuroImage 113:13–25.

    https://doi.org/10.1016/j.neuroimage.2015.03.010

    • PubMed
    • Google Scholar

Author details

1. Xinran Feng

   Department of Psychology & Tsinghua Laboratory of Brain and Intelligence, Tsinghua University, Beijing, China

   Contribution

   Data curation, Formal analysis, Investigation, Visualization, Methodology, Writing - original draft, Writing - review and editing

   Contributed equally with

   Shan Xu

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-4298-1666

2. Shan Xu

   Faculty of Psychology, Beijing Normal University, Beijing, China

   Contribution

   Validation, Investigation, Visualization, Methodology, Writing - original draft, Writing - review and editing

   Contributed equally with

   Xinran Feng

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-6535-3283

3. Yuannan Li

   Department of Psychology & Tsinghua Laboratory of Brain and Intelligence, Tsinghua University, Beijing, China

   Contribution

   Data curation, Investigation, Methodology, Writing - review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0009-0009-6851-4351

4. Jia Liu

   Department of Psychology & Tsinghua Laboratory of Brain and Intelligence, Tsinghua University, Beijing, China

   Contribution

   Conceptualization, Supervision, Funding acquisition, Writing - review and editing

   For correspondence

   liujiathu@tsinghua.edu.cn

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-0383-0934

• 903

  views

• 26

  downloads

• 1

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.