Spatial reorganization of object representations in high-level visual cortex distinguishes working memory from perception

Abstract

The visual system balances veridical object perception with flexible visual working memory (VWM), both engaging high-level visual regions. However, how these competing demands shape object-selective cortex (OSC) representations remains unclear. Here, we ask whether VWM inherits the spatial constraints observed in the OSC during perception or reorganizes representations to meet mnemonic demands. Using a matched task design and functional magnetic resonance imaging–based decoding, we systematically compared the localization of object-identity representations during perception and VWM. A robust dissociation emerged: Perception kept object information largely contralateral, whereas VWM produced robust ipsilateral representation even when memorizing bilateral items. These expanded representations exploited more than 90% of ipsilateral OSC vertices and could not be attributed to increased attentional span. Critically, the ipsilateral representations closely mirrored contralateral representations, implicating interhemispheric coordination in VWM. Together, these findings demonstrate that object VWM flexibly recruits distributed high-level visual cortex, with spatial reorganization distinguishing mnemonic flexibility from perceptual fidelity.

Remembering objects spreads representation across both brain hemispheres, while perception stays spatially local.

INTRODUCTION

The human visual system faces two fundamental challenges: constructing accurate representations of the current environment through perception and maintaining behaviorally relevant information through visual working memory (VWM). Despite their distinct functional goals, perception and VWM often recruit overlapping neural circuits (1–4). Higher-level visual cortex, including the object-selective cortex (OSC), supports object identity encoding during perception (5–7) and is also implicated in maintaining object representations during VWM (8–10). Nevertheless, whether such overlapping recruitment implies that the visual cortex plays a necessary and causal role in VWM remains a matter of ongoing debate (8–12).

Differences in computational demands between perception and VWM may provide a framework for understanding how they engage overlapping regions while supporting distinct functions. Visual perception emphasizes efficient sensory encoding of the external environment, typically through spatially constrained coding schemes. This is reflected in the retinotopic organization of early visual areas and contralateral biases observed in higher visual regions (13–16). In contrast, VWM must maintain internal representations without ongoing sensory input, often in the face of distraction or interference. This functional shift is thought to require more flexible or distributed coding strategies that prioritize stability over spatial specificity (17, 18). These contrasting demands suggest that perception and VWM may impose different spatial constraints on shared cortical substrates, potentially leading to distinct patterns of representational organization.

Evidence from early visual cortex (EVC) suggests that spatial coding is flexible and adapts to task demands. Perceptual representations in low-level visual areas adhere closely to retinotopic organization, with decodable sensory information confined to corresponding retinotopic regions (19). However, during VWM maintenance, this spatial lateralization effect becomes more flexible: Contralateral preferences persist when memorizing bilateral stimuli (20) but notably diminish during unilateral memory maintenance (21, 22). These findings suggest that EVC can shift from a strictly retinotopic code to a more unified mnemonic representation, depending on task context.

This raises an important question: Does similar spatial flexibility extend to higher-level regions such as the OSC, whose computations must reconcile spatial sensitivity with position-tolerant object-level abstraction? The spatial coding properties of OSC are shaped by a nuanced interplay between its anatomical structure, functional tuning, and sensitivity to stimulus configurations (6, 23–25). OSC neurons have large receptive fields that can span both visual hemifields yet exhibit reliable contralateral response biases due to dominant feedforward input from EVC (23, 26). Functionally, OSC supports position-tolerant object recognition: Object identity can be decoded across different locations (27), and category-selective responses (e.g., to faces) are preserved regardless of visual field (6). This position tolerance, however, does not negate contralateral bias—responses are typically stronger for contralateral stimuli than ipsilateral (23, 24). Furthermore, responses in high-level visual areas are modulated by the spatial layout of visual input: Under unilateral presentation, ipsilateral hemisphere shows weaker but detectable representation (28–31), whereas under bilateral viewing, ipsilateral information is strongly suppressed and contralateral stimuli dominate, consistent with interhemispheric normalization (25, 29). Together, these findings illustrate a perceptual regime in which OSC expresses both position tolerance and spatial bias, with the balance shaped by input layout.

While these properties of the OSC are well established during perception, it is less clear whether and how similar spatial coding principles operate during object VWM. Although previous studies proposed the active engagement of OSC during VWM, the nature of its mnemonic representations may differ from its perceptual responses. Converging findings across species suggest that OSC, and more broadly ventral temporal and inferotemporal (IT) cortex, contributes to VWM. Classical neurophysiological studies showed sustained delay-period activity in IT neurons (32), and human functional magnetic resonance imaging (fMRI) work has revealed that category-selective subregions of OSC, such as the fusiform face area (FFA) and parahippocampal place area (PPA), remain active during memory for faces or scenes (33–35). The neural object-file theory further proposes that the posterior parietal cortex provides capacity-limited storage of visual features while the lateral occipital complex encodes content-specific object representations that are modulated by task setting and top-down feedback (36, 37). Together, these findings indicate that high-level visual cortex may play an active role in VWM, yet the representational characteristics, particularly its spatial organization, remain poorly understood.

Given the OSC’s dual sensitivity to spatial position and position-invariant identity, together with its active involvement during VWM, a key question emerges: Does OSC preserve the spatial constraints observed during perception, or does it reorganize to recruit additional ipsilateral resources during object VWM? Moreover, is such flexibility stable across different levels of memory load?

To address these questions, we systematically compared spatial representational patterns in the OSC across perception and VWM using matched bilateral presentation paradigms. Here, we use “spatial representation” to refer specifically to where in the brain object-identity information can be decoded rather than the encoding of stimulus location. Participants viewed or memorized real objects across six conditions: unilateral perception (UP), bilateral perception (BP), central perception (CP), unilateral VWM, bilateral VWM, and bilateral attention (BA) tasks. This design allowed us to examine how the representational structure, particularly the involvement of the ipsilateral hemisphere, varies across tasks.

Our results revealed robust differences between perception and VWM in the spatial representation of objects within the OSC. Specifically, bilateral perceptual configurations result in negligible ipsilateral representation in the OSC, whereas VWM consistently engages ipsilateral regions even when maintaining bilateral items. This VWM-related expansion recruits additional cortical areas beyond those activated during UP or BA tasks and adopts a sensory-like representational pattern. Although both perceptual and mnemonic representations share similar forms, these findings underscore spatial reorganization as a key factor distinguishing perceptual and mnemonic processing in high-level visual cortex.

RESULTS

To directly compare spatial coding in the OSC during perception and VWM, we designed a set of matched fMRI tasks involving unilateral and bilateral object presentation under perceptual, mnemonic, and attentional conditions. Using decoding, representational similarity analysis (RSA), vertex-ablation, and searchlight RSA, we examined how the spatial structure of object representations in OSC flexibly adapts to different task demands. To determine whether these effects were specific to high-level visual areas, we included the EVC as a control region, given its well-characterized retinotopic organization. We also examined the intraparietal sulcus (IPS) and the frontal eye field (FEF) to investigate whether the spatial flexibility of object representations extends beyond the visual cortex to these typically contralaterally biased association areas (38–44). Each participant completed three perception tasks, two VWM tasks, and one attention task under fMRI scanning for more than 11 hours. Regions of interest (ROIs), including the high-level OSC and EVC, were identified using independent localizer scans (see fig. S1, B and C, and Materials and Methods).

Absence of ipsilateral coding in the OSC during bilateral object perception

To establish baseline spatial representations in the OSC during perception, we compared responses to objects presented unilaterally (UP; Fig. 1A) versus bilaterally (BP; Fig. 1B). Twenty objects (fig. S1A) were used across all six tasks. In both UP and BP tasks, participants were required to maintain fixation by performing a fixation-change detection task. We then separately assessed contralateral and ipsilateral representations by decoding object identity using vertices from the ROIs within a single hemisphere. For contralateral representations, for example, we used right-hemisphere vertices to decode objects shown in the left visual field, and vice versa. A 20-class linear support vector machine (SVM) was trained for each decoding analysis, with decoding accuracy calculated via leave-one-trial-out cross-validation (see Materials and Methods).

Fig. 1. Perception task design and results.

(A) Illustration of the UP task. On each trial, a single object was presented on one side of the screen. Participants were required to maintain fixation and respond to any fixation change. (B) Illustration of the BP task. Similar to the UP task, but two objects were presented simultaneously, one on each side of the screen. (C) Decoding accuracy in UP/BP tasks (left) and correlations between representational dissimilarity matrices (RDMs) in UP/BP tasks and RDMs in CP task (right) in OSC. Coral color represents the contralateral condition, while indigo color represents the ipsilateral condition. The horizontal dashed black lines indicate chance level. Error bars represent ±1 SEM across participants. Each symbol corresponds to an individual participant, with consistent symbol-to-participant mapping across all figures (see fig. S5). (D) Same analysis as in (C) but for the EVC. (E) Left: Schematic depiction of OSC object representations during perception. The UP task yielded both contralateral and ipsilateral representations for object identity, whereas the BP task showed object representations only in the contralateral OSC. Right: An open question regarding the neural representation of objects during VWM when sensory input is no longer available, which was further examined in subsequent experiments under comparable encoding conditions. Contra, contralateral; Ipsi, ipsilateral; rep., representation; Dec. acc., decoding accuracy; SEM, standard error of the mean. *P < 0.05, **P < 0.01, and ***P < 0.001; false discovery rate (FDR)–corrected, one-sample t test comparing with chance level. Red asterisks indicate significant differences in ipsilateral representation between the UP and BP task.

In the UP task, decoding accuracy in the OSC significantly exceeded chance level for both contralateral and ipsilateral conditions [Fig. 1C, left; contralateral: t(5) = 3.615, P = 0.015; ipsilateral: t(5) = 2.533, P = 0.026]. In contrast, while contralateral decoding in the BP task remained robust [t(5) = 4.506, P = 0.006], ipsilateral conditions did not reach significance [t(5) = 1.908, P = 0.057]. However, the ipsilateral decoding accuracies did not differ significantly between UP and BP [fig. S2A; t(5) = 2.039, P = 0.068]. All P values in this and subsequent analyses were corrected for multiple comparisons using the false discovery rate (FDR) with a Benjamini-Hochberg procedure (q < 0.05; see details in Materials and Methods and also see tables S1 and S2 for detailed statistics; individual participant data values are listed in table S3).

Because the visual cortex exhibits a contralateral bias, contralateral responses generally have higher signal-to-noise ratios (SNRs) than ipsilateral responses. This SNR asymmetry may cause decoders to fail to detect weaker, yet still present, ipsilateral representations (45). To separate true representational presence from SNR-related limitations, we applied a cross-lateral decoding approach. The decoder was trained on the high SNR contralateral condition and tested on the ipsilateral condition from each task. Consistent with the above ipsilateral decoding results, cross-lateral decoding was significant in the UP [fig. S3A; t(5) = 4.938, P = 0.002] but not in the BP task [t(5) = 1.222, P = 0.138]. Moreover, cross-lateral decoding accuracy was significantly higher in UP than BP task [fig. S2B; UP versus BP; t(5) = 4.997, P = 0.006].

In addition, participants completed another CP task in which a single object was presented at the center of the screen. Responses from this task were used to construct a representational dissimilarity matrix (RDM) template that captures how a unilateral OSC region represents the 20 objects (fig. S4). We then compared each condition’s RDM to this CP-derived RDM template (see details in Materials and Methods). This fixed, sensory-driven reference allows us to assess whether each condition preserves sensory-driven representational format, independent of task-specific training effect in the decoding analysis. Again, ipsilateral representations were only significant in the UP [Fig. 1C, right, and table S4; mean Fisher Z = 0.328, t(5) = 3.610, P = 0.008] but not in the BP task [mean Fisher Z = 0.035, t(5) = 0.907, P = 0.203], and the UP ipsilateral representation is stronger than BP [fig. S2C and table S2; t(5) = 4.888, P = 0.003], indicating an absence of sensory-based ipsilateral coding under bilateral visual input. Meanwhile, there was no difference in the noise ceiling levels between the two tasks (fig. S5). Binary decoding analyses for each object pair and the decoding time course based on the blood oxygen level–dependent (BOLD) signal also yielded similar results (figs. S6 and S7).

In the EVC (Fig. 1D), both decoding and RSA consistently revealed robust contralateral representation in the UP and BP tasks (all P values < 0.001). By contrast, ipsilateral decoding and RSA results failed to reach significance in either the UP (all P values ≥ 0.069) or the BP task (all P values > 0.100). These findings suggest that ipsilateral representations in EVC are minimal and remain unaffected by changes in input configuration. This pattern is consistent with the strict retinotopic organization of EVC (19, 46).

Collectively, these results demonstrate that ipsilateral representations in the OSC are robust under unilateral presentation but largely absent under bilateral conditions, consistent with previous findings (25, 29). This reflects an adaptive spatial organization modulated by the visual input configuration, providing a baseline for the subsequent comparison with the VWM representation.

VWM recruits bilateral OSC regions for both single and bilateral object memory

Given the absence of ipsilateral OSC representations during BP (Fig. 1E, left), we next asked whether similar spatial constraints govern VWM (Fig. 1E, right). If VWM maintains the same perceptual representations, then memorized object identity should remain lateralized. Alternatively, VWM might recruit additional ipsilateral areas to support a more distributed and bilateral memory code.

To test these two competing hypotheses, we designed two VWM tasks that differed in memory load. In the single-item postcue WM task (one-item WM), two objects were presented bilaterally, followed by a spatial cue indicating which side’s object participants should maintain in memory (Fig. 2A). Participants performed this task with high accuracy (mean accuracy = 0.955, SEM = 0.012; fig. S8A), indicating strong task engagement.

Fig. 2. VWM task design and results.

(A) Illustration of the one-item postcue WM task. Participants maintained fixation throughout the trial and memorized the cued object during the delay period. Afterward, they reported both the presence and identity of the cued object. (B) Illustration of the two-item postcue WM task. Similar to (A), but participants had to retain both objects during the delay, as the cue was presented after the delay period. (C) Schematic of decoding analysis using single-hemisphere ROIs. Taking the contralateral cued condition as an example (left diagram), for trials where the cued object was presented in the right visual field, only the BOLD signals from the left hemisphere (LH) ROIs were used for the decoding analysis (see Materials and Methods for details). (D) Decoding accuracy under contralateral (left) and ipsilateral (right) conditions for the one-item postcue WM task. Line charts show the time course of decoding accuracy calculated using BOLD signals at each TR (repetition time; TR = 1 s), while bar charts represent decoding accuracies computed using averaged BOLD signals across 6 to 10 TRs. In the bar charts, empty and dashed bars indicate cued and uncued conditions, respectively. (E) Decoding accuracy under contralateral (left) and ipsilateral (right) conditions for the two-item postcue WM task. The horizontal dashed black lines indicate chance level. Coral color represents the contralateral condition, while indigo color represents the ipsilateral condition. Shaded areas and error bars represent ±1 SEM. *P < 0.05, **P < 0.01, ***P < 0.001, and n.s. (not significant); FDR-corrected, one-sample or paired t test. Red asterisks indicate significance in the paired comparison between the current task’s ipsilateral representation and that of the BP task.

We performed decoding analyses using BOLD signals at each TR (repetition time), separating cued and uncued objects under contralateral and ipsilateral conditions (Fig. 2C). As an example of this procedure, in the contralateral-cued condition, we decoded object identity using the ROI vertices in the opposite hemisphere. The same procedure was applied for ipsilateral-cued trials using the corresponding hemisphere (see Materials and Methods). Results from left- and right-field trials were averaged to generate the overall decoding time course (Fig. 2D). For statistical comparisons, we also computed decoding accuracy based on the mean BOLD activity during the delay period (TRs of 6 to 10; see fig. S8B for deconvolved responses and fig. S9 for results using later time windows).

We first conducted two-way repeated-measures analyses of variance (ANOVAs) to assess laterality (contralateral versus ipsilateral) and cueing (cued versus uncued) effects. The analysis showed a significant interaction between laterality and cueing [F(1,5) = 12.045, P = 0.018]. The main effects of laterality [F(1,5) = 15.325, P = 0.011] and cueing [F(1,5) = 16.553, P = 0.009] were also significant. As shown in the left bar chart of Fig. 2D, contralateral decoding was significantly stronger for cued compared to uncued objects [t(5) = 4.870, P = 0.004], and both cued and uncued objects were decoded above chance (both P values < 0.01; see tables S5 and S6; individual participant data values are listed in table S7). In the ipsilateral hemisphere (Fig. 2D, right), the cueing effect still existed, showing higher decoding accuracy for cued than uncued objects [t(5) = 2.854, P = 0.027]. However, significant decoding was observed exclusively for cued objects [t(5) = 3.651, P = 0.022], whereas decoding accuracy for uncued objects remained at chance [t(5) < −0.001, P = 0.500]. Similar results were obtained with cross-lateral decoding (fig. S3B) and binary decoding (fig. S10, left). These results confirmed the critical involvement of OSC in object VWM.

In the EVC, the interaction between the laterality and cueing was significant [F(1,5) = 10.486, P = 0.023]. There was also a significant main effect of laterality [F(1,5) = 10.210, P = 0.024], whereas the main effect of cueing was not significant [F(1,5) = 0.952, P = 0.374]. However, contralateral decoding did not differ between cued and uncued objects [t(5) = 1.831, P = 0.063], although both exceeded chance [cued: t(5) = 3.316, P = 0.032; uncued: t(5) = 2.431, P = 0.044]. In the ipsilateral hemisphere, neither cued nor uncued objects were decodable above chance, and no cueing effect was observed (all P values ≥ 0.621). This also held true using cross-lateral decoding (all P values ≥ 0.603). The absence of cueing effect or ipsilateral representation of objects in EVC likely reflects the task’s emphasis on object identity rather than fine-grained visual features (47), which typically drives stronger involvement of early visual areas in VWM (2, 48).

Furthermore, to directly compare the ipsilateral OSC representations during perception and VWM, we compared the ipsilateral decoding accuracy between the VWM and BP task. The results showed significantly stronger ipsilateral representation in the one-item WM than in the BP task using both ipsilateral decoding [fig. S2A; t(5) = 3.570, P = 0.028] and cross-lateral decoding [t(5) = 4.443, P = 0.006]. These findings reveal that, unlike the contralateral-only representation observed in the BP task, object VWM recruits both contralateral and ipsilateral OSC. In contrast, neither hemisphere of EVC supports VWM representations. This suggests that the OSC plays a key role in object VWM, with memory representations expanding to both hemispheres rather than being strictly constrained by perceptual configurations.

Would ipsilateral memory representations still emerge when participants must remember two objects presented in both visual fields? To address this question, we used a two-item postcue WM (two-item WM) task in which the spatial cue was presented only at the end of the trial (Fig. 2B). In this modified setting, participants needed to retain both objects and their respective locations throughout the delay. Despite the increased memory load, participants still achieved high accuracy (mean accuracy = 0.947, SEM = 0.008; fig. S8A). If ipsilateral decoding in OSC failed under this condition, it would suggest that such representations emerge only when a single object is prioritized, indicating task-dependent lateralization. Alternatively, successful decoding would support the notion that bilateral representations are a general property of object VWM.

In the two-item WM task, object identity remained decodable in the contralateral OSC [Fig. 2E, left; t(5) = 5.240, P = 0.003; see table S8; individual participant data values are listed in table S9]. The representations were also successfully decoded under ipsilateral conditions [Fig. 2E, right; ipsilateral decoding: t(5) = 4.555, P = 0.003] [fig. S3B, right; cross-lateral decoding: t(5) = 5.656, P = 0.001]. Consistently, ipsilateral EVC showed no significant decoding [ipsilateral decoding: t(5) = 0.411, P = 0.349; cross-lateral decoding: t(5) = 1.247, P = 0.134]. Moreover, the two-item WM ipsilateral decoding was also stronger than in the BP task [ipsilateral decoding (fig. S2A): t(5) = 3.927, P = 0.028; cross-lateral decoding (fig. S2B): t(5) = 7.019, P = 0.004], consistently showing an extended ipsilateral mnemonic representation than in the perceptual condition.

These results demonstrate that, unlike perception, VWM representations in the OSC are not strictly constrained by spatial lateralization. Instead, object information extends to ipsilateral regions, reflecting a more flexible coding scheme. In addition, this ipsilateral recruitment persists in the two-item condition, indicating that bilateral representation is a general feature of object VWM in the OSC.

VWM extends ipsilateral OSC recruitment beyond perceptual constraints

Our previous analysis revealed that, although ipsilateral representation is minimal under bilateral input, VWM engages a robust ipsilateral representation, even under increased memory demands. However, the characteristics underlying this expansion remain unclear: (i) Do the additional ipsilateral regions recruited during VWM correspond to those that encode ipsilateral stimuli during perception, as in the UP task? (ii) Does the extent of ipsilateral recruitment vary as a function of memory demand?

To address these questions, we conducted a vertex-ablation decoding analysis. First, we quantified each OSC vertex’s bilateral responsivity by measuring the similarity between its mean responses to 20 objects under contralateral and ipsilateral conditions in the UP task. Vertices with high bilateral responsivity were assumed to have a bilateral receptive field and to be inherently responsive to ipsilateral input during perception. We then ranked all vertices by this metric and progressively removed those with the highest values. After each removal step, we repeated the decoding analysis on the remaining (more unilateral) vertices. The resulting drop in decoding accuracy indicates the fraction of OSC vertices that support ipsilateral VWM representations (Fig. 3A).

Fig. 3. Vertex-ablation analysis in perception and VWM tasks.

(A) Schematic of the vertex-ablation analysis. For each OSC vertex, we calculated bilateral responsivity as the correlation between beta values from ipsilateral and contralateral conditions in the UP task. Vertices with the highest bilateral responsivity were progressively removed. (B) Decoding accuracy in OSC following stepwise vertex-ablation analysis for the perception (first row) and VWM (second row) tasks. Each point along the curve represents decoding accuracy after excluding the top N% of vertices with the highest bilateral responsivity. For the one-item WM tasks, vertex-ablation analysis was conducted only on the cued condition. Horizontal colored lines above each figure indicate ablation steps where decoding was significant above chance (P < 0.05; FDR-corrected across 12 ablation steps, one-sample t test). The horizontal dashed black lines indicate chance level. Coral and indigo represent the contralateral and ipsilateral conditions, respectively. Shaded areas represent ±1 SEM.

The ablation decoding results (Fig. 3B) demonstrated that stable decoding of ipsilateral object information in the one-item WM task persisted even after removing 95% of the cortical vertices (P values < 0.05; corrected across all ablation steps; the same correction was applied for all vertex-ablation analysis for each task; individual participant data values are listed in table S10), with 40 to 80% of the vertices showing above-zero bilateral responsivity across participants (fig. S11). In the two-item WM task, significant decoding remained after removing 90% of the OSC vertices (P values < 0.05; see fig. S12A for results using later time windows). In contrast, in either perception task, no significant ipsilateral decoding was found in any steps. This dissociation remained even under cross-lateral decoding: In the UP task, ipsilateral object identity became undecodable after 60% of vertices were removed, whereas both VWM tasks continued to support significant decoding even after removing more than 95% of vertices (fig. S3C).

These findings indicate that ipsilateral representations of objects in VWM expand to regions beyond what is observed in perception. VWM involves not only vertices inherently responsive to bilateral input but also much of the ipsilateral OSC.

Increased attentional span plays a limited role in ipsilateral expansion

A potential confound is that the difference in attentional span between VWM and perception tasks may have contributed to our results, as an extended attentional range could also alter receptive field properties in the OSC (49). Could the extended ipsilateral representation observed in VWM simply reflect a wider attentional range during encoding compared to the passive viewing condition in the perception task?

To address this concern, we conducted a BA task (Fig. 4A), where one of the objects rotated by a variable angle at random intervals. Participants pressed a key upon detecting a rotation in either visual field while maintaining central fixation. To ensure high attentional engagement, we involved 90% of trials in a rotation, and the rotation angle was adjusted via the QUEST procedure to ensure that overall accuracy converged to 70% (50).

Fig. 4. Attention task design and results.

(A) Illustration of the BA task. Similar to the BP task, two stimuli were presented three times in quick succession in each trial. Participants were required to detect the rotation of objects via a key press. (B) Decoding accuracy for the contralateral and ipsilateral conditions in the OSC and EVC. (C) Vertex-ablation decoding results for the BA task, analogous to Fig. 3. Horizontal colored lines above indicate ablation steps where decoding was significant above chance (P < 0.05; FDR-corrected across 12 ablation steps, one-sample t test). The horizontal dashed black lines indicate chance level. Coral and indigo represent the contralateral and ipsilateral conditions, respectively. Shaded areas represent ±1 SEM. *P < 0.05, **P < 0.01, and ***P < 0.001; FDR-corrected, one-sample t test comparing with chance level.

Decoding analyses showed above-chance decoding accuracy in the ipsilateral OSC [Fig. 4B; ipsilateral decoding: t(5) = 2.421, P = 0.030; see table S11; individual participant data values are listed in table S12], with similar findings from cross-lateral decoding, binary decoding, and decoding time course based on the BOLD signal (figs. S3A, S6, C and D, right, and S7). However, vertex-ablation decoding indicated that decoding did not remain significant at any ablation level after correction for multiple comparisons (Fig. 4C). Under cross-lateral decoding, ipsilateral object identity became undecodable after 70% of vertices were removed (fig. S3C), still reflecting a more spatially constrained ipsilateral representation compared to the VWM tasks.

Direct task comparisons further underscored the greater expansion of ipsilateral representations during VWM. Although the BA task showed stronger ipsilateral decoding than the BP perception task in cross-lateral decoding [ipsilateral decoding (fig. S2A): t(5) = 0.881, P = 0.209] [cross-lateral decoding (fig. S2B): t(5) = 3.693, P = 0.008], ipsilateral decoding in both VWM tasks was stronger than in the BA task across decoding approaches [fig. S2A; ipsilateral decoding: one-item WM versus BA, t(5) = 2.550, P = 0.047; two-item WM versus BA, t(5) = 2.509, P = 0.047] [fig. S2B; cross-lateral decoding: one-item WM versus BA, t(5) = 4.192, P = 0.006; two-item WM versus BA, t(5) = 4.575, P = 0.006]. Together, these results indicate that while an extended attentional span during encoding may contribute to ipsilateral recruitment, it cannot fully account for the broader ipsilateral representation observed in VWM.

Cortical distribution of sensory-based ipsilateral representation

In the above ROI decoding analyses, we found that object VWM engages both contralateral and ipsilateral OSC, with vertex-ablation decoding further revealing substantially more ipsilateral vertices contributing to VWM than to perception or attention. These findings indicate that VWM relies on an expanded spatial neural substrate. However, while decoding reveals the presence of ipsilateral information during VWM, it does not characterize the underlying representational format. We therefore used RSA to test whether ipsilateral VWM activity shares the sensory-based format measured during perception.

Following the same approach used in the perception tasks, we computed ROI-based RDMs for the attention and VWM tasks and compared them with the CP-derived sensory template RDM. Across all tasks, contralateral OSC showed robust correlations with this sensory template (fig. S13A; all P values < 0.005, corrected for multiple comparisons across five tasks), demonstrating a stable sensory-like format. Critically, ipsilateral OSC exhibited significant sensory-based representation only during VWM [one-item WM: t(5) = 6.181, P = 0.005; two-item WM: t(5) = 5.134, P = 0.005], whereas no such correlations were observed in the BP and BA tasks [BP: t(5) = 0.907, P = 0.203; BA: t(5) = 1.612, P = 0.105]. Ipsilateral representations were stronger in both VWM tasks than in the BP and BA (fig. S2C; all P values < 0.003) tasks and stronger in one-item than two-item WM [t(5) = 3.092, P = 0.016]. Similar patterns were also observed using RSA on the BOLD signal time courses (fig. S13B). These RSA findings largely mirrored the decoding results (fig. S13C), indicating that VWM reinstates a sensory-like representational format even for the extended ipsilateral activity.

Having established the representational format, we next examined where these sensory-based ipsilateral representations are expressed. To visualize the cortical distribution of these representations across tasks, we used an RSA-based searchlight analysis (Fig. 5). Each participant’s OSC RDM template, as used above, served as a common reference. For each task, we constructed an RDM for each vertex based on its neighboring vertices and correlated these RDMs with the participant’s reference RDM. The resulting RSA maps were projected onto the common surface (fsaverage; Fig. 5 and fig. S14; see fig. S12B for results using later time windows). This analysis was applied to two VWM tasks (cued condition in one-item WM and two-item WM), as well as the UP, BP, and BA tasks.

Fig. 5. Cortical distribution of sensory-based ipsilateral representations.

(A) Schematic depiction of searchlight RSA. The RDM from the CP task in the OSC served as the reference RDM (fig. S4) and was correlated with floating RDMs from spatially localized searchlights in different tasks (see Materials and Methods). (B) Ipsilateral representations revealed by searchlight RSA across perception, attention, and VWM tasks. Only vertices with Pearson r values > 0.1 are shown. “LV” (left visual field) and “RV” (right visual field) labels indicate that values on each hemisphere were calculated on the basis of the stimulus label in the corresponding visual field. For the LH, values correspond to the LH-LV subcondition (LH activation for LV stimuli); for the right hemisphere (RH), values correspond to the RH-RV subcondition (RH activation for RV stimuli).

Focusing on cortical distribution of the representation to the object presented in the same-side visual field, we observed strong ipsilateral representations during the UP task but only limited ipsilateral engagement in the BP task [Fig. 5B and fig. S15; averaged vertices correlation versus zero; UP: t(5) = 4.381, P = 0.008, BP: t(5) = 0.627, P = 0.279]. This was supported by higher averaged correlation values among OSC vertices in the UP compared to BP [fig. S2D; t(5) = 4.492, P = 0.007; see table S2].

In VWM, the one-item WM task exhibited a broader ipsilateral representation than in the BP and BA tasks [fig. S15; one-item WM versus zero: t(5) = 5.727, P = 0.005] [fig. S2D; one-item WM cued versus BP: t(5) = 4.527, P = 0.007; one-item WM cued versus BA: t(5) = 5.918, P = 0.007]. The two-item WM task also showed broader sensory-based ipsilateral representation [fig. S15; two-item WM versus zero: t(5) = 4.042, P = 0.008], as evidenced by higher correlations than those in BP [t(5) = 3.174, P = 0.022] and BA task [t(5) = 2.506, P = 0.038]. The extent of ipsilateral sensory representation did not differ significantly between the one-item and the two-item WM tasks [t(5) = 1.623, P = 0.083].

Together, our results show that object VWM is associated with both the expansion and spatial redistribution of sensory-like ipsilateral coding across the OSC. This finding indicates a clear reorganization of representational topology beyond what is captured by global ROI-based RSA and ablation metrics alone.

Spatially organized association regions show distinct contributions to object VWM

In addition to high-level visual regions, association areas, such as IPS and frontal regions, may also play a critical role during VWM. Delay-period activity and stimulus-related mnemonic representations have been reported in these regions (1, 2, 36, 51–57). Certain parietal and frontal regions exhibit spatially organized, contralaterally biased topographic maps (38–40). This spatial organization is preserved during VWM, with the IPS showing contralateral mnemonic selectivity (58, 59). Moreover, converging evidence indicates that the same retinotopic spatial-priority maps are engaged across perception, attention, and VWM in both IPS and the human FEF (41–44). Electrophysiological recordings in nonhuman primates further demonstrate that VWM traces in frontal cortex are maintained in a contralateral, retinotopic reference frame and are transferred across hemispheres when gaze shifts (60). Together, these findings establish that spatial constraints on visual perception and VWM extend beyond sensory cortex into association networks.

We therefore asked whether these spatially constrained association areas would also exhibit the bilateral expansion of object identity representations that we observed in the OSC. To test for this, we conducted decoding analysis in the IPS and FEF in our VWM tasks. In the IPS, we found a significant interaction between laterality and cueing in the one-item WM task [F(1,5) = 6.882, P = 0.047]. The main effects of laterality [F(1,5) = 7.642, P = 0.040] and cueing [F(1,5) = 56.510, P < 0.001] were also significant. As shown in Fig. 6, contralateral IPS showed robust cueing effect, displaying significantly higher decoding accuracy for cued compared to uncued objects [t(5) = 6.841, P < 0.001], and only the cued objects showed significant above-chance decoding [cued: t(5) = 7.559, P < 0.001; uncued: t(5) = −1.423, P = 0.893; see tables S5 and S6; individual participant data values are listed in table S7]. The cueing effect was also observed in ipsilateral IPS [t(5) = 4.890, P = 0.003], with significant above-chance decoding only for cued items [cued: t(5) = 6.181, P = 0.002; uncued: t(5) = −1.211, P = 0.860]. This bilateral representation persisted in the two-item WM task, showing both significant decoding under contralateral [t(5) = 6.845, P = 0.001; see table S8; individual participant data values are listed in table S9] and ipsilateral [t(5) = 3.260, P = 0.011] conditions. These results suggest that IPS supports a more global, object-based VWM code as in the OSC.

Fig. 6. Decoding accuracy in association areas in VWM tasks.

Decoding accuracy in the IPS and FEF during one-item WM (left) and two-item WM (right) tasks. (Top) Contralateral decoding results. (Bottom) Ipsilateral decoding results. The horizontal dashed black lines indicate chance level. Error bars represent ±1 SEM. *P < 0.05, **P < 0.01, ***P < 0.001, and n.s. (not significant); FDR-corrected, one-sample or paired t test.

Having established that IPS, similar to OSC, supports bilateral object decoding during WM, we next asked whether IPS also preserves the sensory-like representational format for the bilateral code. To test this, we applied the same ROI-based RSA approach used for the OSC, comparing IPS RDMs during VWM with the CP-derived sensory template RDM. Unlike the OSC, where both contralateral and ipsilateral VWM representations showed significant correlations with the sensory template (fig. S13A), IPS showed no significant correlation with the CP template in either hemisphere or task (fig. S16A; all P values > 0.570). Time-resolved RSA confirmed this null result, with no significant time points observed in IPS across any condition (fig. S16B).

To directly quantify the regional dissociation, we conducted a two-way repeated-measures ANOVA (region × laterality) on the RSA results. In both VWM tasks, this revealed a significant main effect of region [one-item WM: F(1,5) = 33.790, P = 0.002; two-item WM: F(1,5) = 9.902, P = 0.025] and a significant interaction [one-item WM: F(1,5) = 9.667, P = 0.026; two-item WM: F(1,5) = 11.462, P = 0.019]. Follow-up simple-effect analyses confirmed that the sensory representational format was significantly stronger in OSC than in IPS for both the contralateral [one-item WM: t(5) = 6.074, P = 0.001; two-item WM: t(5) = 3.265, P = 0.011] and ipsilateral conditions [one-item WM: t(5) = 5.312, P = 0.008; two-item WM: t(5) = 2.904, P = 0.028]. These findings indicate that, although IPS supports bilateral object decoding during VWM, its representational format does not preserve the sensory-like geometry observed in the OSC.

In contrast, we did not observe reliable evidence for object identity decoding in FEF under either contralateral or ipsilateral conditions in the one-item WM task. The interaction between laterality and cueing effect, as well as the main effects, was not significant (Fig. 6; all P values > 0.213). Neither the cued nor uncued items showed above-chance decoding under contralateral or ipsilateral conditions (all P values ≥ 0.201; see tables S5 and S6; individual participant data values are listed in table S7). Similarly, in the two-item WM task, neither contralateral nor ipsilateral representations were significantly decodable (all P values > 0.354; see table S8; individual participant data values are listed in table S9). These results indicate that, under our task conditions and measurement resolution, we did not detect robust item-level object representations in FEF during VWM. This does not imply that FEF does not contribute to object VWM. Prior work suggests that FEF may contribute to the control, prioritization, and spatial organization of mnemonic contents (43, 61, 62). These control-related signals, as well as potentially more abstract or nonlinear representational formats, may not be detectable using standard multivoxel decoding approaches. In addition, the absence of detectable multivoxel patterns may reflect the spatial resolution and temporal smoothing inherent to fMRI rather than a true absence of mnemonic coding.

We next asked whether association areas share representational pattern with the ipsilateral OSC. To test this, we performed cross-region RSA for both VWM tasks, comparing the representational patterns of ipsilateral OSC with contralateral OSC and other regions. Using partial regression to isolate the unique contributions of each region’s RDM under each condition, we found that ipsilateral OSC representations correlated most strongly with contralateral OSC, more so than with IPS or FEF (fig. S17). This correlational analysis indicates that ipsilateral representations in the OSC likely mirror contralateral patterns, suggesting potential interhemispheric interactions. Although we did not observe a strong correlation between ipsilateral OSC and frontal and parietal areas, this does not rule out the possibility of top-down contributions, as feedback signals may undergo representational transformation along the visual hierarchy (45, 54).

Collectively, these results echo previous findings highlighting the involvement of association areas in maintaining mnemonic information (1, 51–54, 57). Furthermore, ipsilateral OSC more closely exhibited a representational structure that most resembled contralateral OSC, suggesting that the bilateralization observed in OSC reflects a content-preserving extension of its own representational format. By contrast, the absence of sensory-template correlation in IPS indicates that ventral and parietal regions contribute to object VWM through distinct representational mechanisms, with OSC preserving perceptual coding and IPS supporting a more abstract or task-driven format (53, 54, 63).

DISCUSSION

The present study reveals a stable difference in the spatial organization of object representations between visual perception and VWM within the OSC. While perceptual representations in the OSC are spatially constrained by visual field occupancy, VWM recruits ipsilateral cortical regions beyond those engaged during perception and attention. Moreover, this expanded representation persisted across varying memory demands, and ipsilateral OSC representations closely aligned with contralateral OSC. Bilateral IPS also contributes to memory maintenance, but only contralateral IPS shows significant similarity with contralateral OSC. This pattern suggests that the expanded ipsilateral OSC retains a sensory-like representational format, potentially supported by interhemispheric interactions, while top-down feedback from association areas may be transformed. Together, these findings highlight the existence of spatial reorganization in high-level visual cortex during VWM compared to perception, revealing a more flexible and globally distributed coding strategy for VWM.

This reorganization refines prior accounts of how OSC participates in VWM. Previous studies have demonstrated that OSC retains stimulus-specific information during VWM and preserves aspects of category selectivity, a hallmark of object perception (35, 64). However, it has remained unclear how the more goal-directed nature of VWM alters OSC representations relative to perception. By demonstrating that VWM reorganizes, rather than simply reinstates, perceptual codes, our results extend the object-file framework (36, 37). While this framework posits a division between a spatial indexing system in IPS and a content-based subsystem in OSC, later work has shown that VWM representations overlap with perceptual regions, and their format can shift with task demands and attention (1, 3, 57). Extending this view, our results show that OSC not only preserves perceptual content but also adapts its spatial topology to support internal representations, highlighting a dynamic transition from perceptual to mnemonic states tailored to behavioral demands. Our contribution also extends the influential framework that the locus of VWM storage varies with stimulus type (17). While this framework addresses where VWM representations reside, our results reveal how they are spatially and representationally organized within a given region (65–67). Specifically, the combination of intraregional spatial reconfiguration and preservation of sensory format we observed in OSC identifies representational properties for object VWM that cannot be inferred from the locus of storage alone.

These findings also speak to the broader debate about whether the sensory visual cortex is necessary for VWM (8–12, 65–75). While many studies have decoded mnemonic content from visual cortex (1–4, 20, 48, 55, 56, 76, 77), other work suggests that these signals may be transient, susceptible to distraction, with sustained storage attributed to parietal or frontal regions (51, 53, 54, 57, 78–81). Although our study does not resolve this question, it offers a different perspective from the high-level visual cortex that VWM alters the spatial extent of representational formats, relaxing perceptual biases and enabling bilateral recruitment. This dynamic spatial reconfiguration echoes some of the flexibility previously observed in EVC (21, 22), suggesting that such transformations may be a more general property of the visual system under mnemonic load. Together, our findings align with the view that advancing this field requires not only identifying where memory representations emerge but also understanding how these representations transform across sensory and association areas and which factors regulate their maintenance during VWM (17).

Our results further clarify how perception and VWM differ in OSC. Although OSC neurons have large receptive fields supporting position-tolerant object coding, they also exhibit a reliable contralateral bias shaped by competitive interactions during bilateral input (25, 29, 82). Our findings are consistent with this view and further demonstrate that, during bilateral object presentation, object coding in OSC is largely confined to the contralateral hemisphere, likely reflecting competitive interactions that prioritize contralateral input. We also find that attention can modulate this spatial pattern, weakening contralateral dominance when increasing attention is distributed more evenly to both objects. However, this spatial constraint is largely overcome during VWM, where robust bilateral representations persist even when maintaining two objects simultaneously, exceeding the ipsilateral recruitment observed in perception and attention tasks. These findings demonstrate that the distinction between perception and working memory in OSC does not reflect a categorical change in representational code but rather a reorganization of spatial topology that preserves sensory representational geometry. This spatial expansion with format preservation is not a generic property of all regions showing ipsilateral WM signals. IPS also exhibits significant bilateral decoding during VWM, but IPS representations show no significant correlation with the sensory template derived from CP, indicating that its ipsilateral WM signals reflect a more abstract or task-driven code rather than a reinstatement of perceptual format (53, 54, 63, 83). This dissociation suggests that ventral and parietal cortices contribute to object VWM through complementary yet mechanistically distinct strategies: OSC reconfigures its spatial coding while retaining sensory content, whereas IPS supports mnemonic representations in a format that departs from perceptual coding.

During perception, spatial competition and attention jointly shape contralateral biases, producing highly localized, contralaterally biased representations. During VWM, when sensory input is absent, this spatial constraint relaxes, and the same cortical territory adopts a more distributed organization that extends into ipsilateral regions. This adaptive widening of the representational field may support the sustained maintenance of object identity when bottom-up input is no longer available.

A finding of this study is that VWM-related spatial expansion is selective to OSC rather than universally distributed across the visual hierarchy. While EVC can exhibit ipsilateral representations when retaining low-level features (20–22), we did not observe robust ipsilateral EVC decoding in our VWM tasks. We caution that this could reflect the resolution of mnemonic demands in our paradigm. In the current task, the participants were required to memorize the object identity at the basic category level (e.g., “elephant”) rather than at the exemplar level (e.g., “which particular elephant”), which does not require maintaining fine-grained visual detail that is typically associated with robust recruitment of EVC (2, 48). Hence, the current null results at EVC may be explained by reliance on a coarser visual code rather than providing strong evidence that object VWM does not occur in EVC (47). This leaves an important question: Is memory-related spatial expansion driven by stimulus complexity or task relevance (20, 22, 84)? Future studies should systematically manipulate the level of visual detail required during memory (e.g., exemplar-level discrimination) to determine whether VWM for real-world objects engages EVC in a similar bilateral manner.

Understanding the mechanisms underlying the expanded spatial representation in VWM remains a critical question. One possibility is that this expansion arises from changes in population receptive field (pRF) properties during encoding. VWM requires a broader attentional span than perception, which could extend receptive fields and allow ipsilateral encoding (49). However, our data suggest that such attentional modulation alone is insufficient to account for the broad ipsilateral representations observed in VWM (Fig. 4), implying additional underlying mechanisms. Another possibility is that ipsilateral representations arise via top-down feedback from association cortices such as the parietal and frontal regions (21). Classic neurophysiological work has demonstrated that prefrontal cortex maintains stimulus-selective delay activity and can bias representations in ventral visual cortex through feedback (85), suggesting that top-down input could indeed support VWM-related signals in OSC. However, our cross-region RSA suggests that ipsilateral OSC does not simply adopt the representational geometry of IPS or FEF, as its representational structure more closely aligned with contralateral OSC (fig. S17). Because feedback signals are known to transform as they propagate through the hierarchy (54), this result neither confirms nor rules out top-down influence but indicates that the expanded mnemonic representation in ipsilateral OSC retains a ventral-stream format. A further possibility is that intrinsic visual network dynamics, particularly interhemispheric coordination and communication between homologous OSC regions, may drive this spatial expansion (21, 86). While our cross-RSA analysis aligns with this hypothesis (fig. S17), the limited temporal resolution of fMRI prevents direct causal inferences. Future studies could address this by leveraging high–temporal-resolution methods, such as high-density electrophysiological recordings or intracranial electroencephalography (EEG), to simultaneously record bilateral IT cortex activity during memory tasks in primates, thereby clarifying the mechanisms of interhemispheric information communication.

An additional methodological limitation concerns fixation monitoring. Although participants were instructed to maintain central fixation and performed a fixation-change detection task during perception, we did not collect in-scanner eye-tracking data. The sample arrays were presented briefly (200 ms) to minimize opportunities for saccades, and critically, EVC exhibited strictly contralateral decoding across all tasks, which is a pattern that would likely be disrupted by systematic gaze shifts. Even so, a more systematic characterization of eye-movement behavior across perception, attention, and VWM tasks would be valuable for understanding how oculomotor and mnemonic systems jointly shape spatial coding. Incorporating high-resolution gaze monitoring in future work would not only help confirm fixation stability but also offer the opportunity to examine whether fine-grained oculomotor dynamics themselves participate in the reorganization of representational topology.

A further consideration concerns the temporal separation of perceptual and mnemonic signals. Because the primary analysis window (TR of 6 to 10) follows stimulus onset by only a few seconds, it may retain residual stimulus-evoked responses, as suggested by above-chance contralateral EVC decoding for both cued and uncued items during this period (fig. S9). Critically, however, perceptual contamination cannot account for the ipsilateral OSC representations reported here. In the one-item WM task, cued and uncued objects shared identical perceptual input during encoding, yet ipsilateral OSC decoding was significant only for the cued item, with the uncued item remaining at chance (Fig. 2D), as confirmed by a significant laterality × cueing interaction. This cue-specific pattern is inconsistent with a purely stimulus-driven account. Moreover, analyses using later time windows, in which contralateral EVC decoding returned to chance, continued to show significant ipsilateral OSC decoding (fig. S9) and yielded consistent vertex-ablation and searchlight RSA results (fig. S12). Although the limited temporal resolution of fMRI precludes a complete separation of encoding-related and maintenance-related signals (72), the selective association of ipsilateral OSC representations with mnemonic status, together with their persistence beyond the perceptual response window, provides converging evidence for a memory-related rather than perceptual origin.

While these analyses support a mnemonic interpretation of the ipsilateral expansion, they raise a further question: When during the trial does this bilateral reorganization emerge? While fMRI cannot resolve these temporal dynamics, EEG studies provide relevant insights. Previous studies have shown a strong contralateral bias during encoding but equally strong decoding from bilateral posterior channels during VWM reactivation (87), suggesting a post–encoding shift toward bilateral memory traces over time. A relevant phenomenon that may be related to the emerging reorganization comes from the contralateral delay activity (CDA) in EEG. The difference between contralateral and ipsilateral responses relative to the visual field of the remembered items reflects VWM storage (88). The contralateral-ipsilateral differences gradually reduced during the delay period, as observed in both CDA amplitude reductions (89), and declines in the decoding accuracy based on the bilateral response difference (90). These findings suggest that VWM reorganizes object representations spatially, shifting from contralateral dominance to bilateral encoding during the delay period. However, it remains unclear whether this shift reflects a weakening of the strong contralateral sensory input or the construction of a new memory trace, as well as where in the cortex this reorganization takes place, given the spatial limitations of EEG. Further computational modeling and precise neural recordings are needed to resolve this question.

Although we observed spatially expanded representations in VWM, other studies have reported task-dependent spatial constraints on memory representations. In early visual areas, bilateral representations are observed when a single item is remembered without a location requirement (22), but when both visual fields are task relevant, representations become more spatially constrained (20). Behavioral studies support this task-dependent flexibility, showing that when participants encode orientations across both visual fields, the stored feature can modulate perception in a spatially specific manner. However, when only one grating is remembered, its influence extends more globally, unless spatial specificity is reinforced (84). In our study, the two-item WM task required participants to retain both the identity and location of the objects, as the spatial cue was presented only after the delay period. Despite this requirement, both our one-item and two-item WM tasks consistently revealed bilateral representations. While maintaining categorical object identity may differ from the high-precision feature requirements in previous EVC experiments (20–22), our findings indicate that the need to maintain spatial information does not preclude the emergence of a globally distributed, bilateral memory code in the OSC.

In summary, our findings reveal a spatially extended, sensory-based VWM representation in the OSC. Unlike perception, which is constrained by spatial organization, VWM flexibly expands its representational range based on input configuration, task modality, and memory demands. Future studies should use EEG and magnetoencephalography to pinpoint the precise temporal dynamics of bilateral representation emergence in VWM. In addition, high-throughput electrophysiological recordings in nonhuman primates could provide high–temporal-resolution insights into interhemispheric coordination in object-selective regions at the single-unit level. Comparing different probing methods may also clarify their influence on VWM representational structure. Together, these approaches will advance our understanding of the neural mechanisms driving the spatial expansion of VWM and its interactions with perception, attention, and executive function.

MATERIALS AND METHODS

Participants

Six healthy participants (three males, age range = 21 to 29 years, mean age = 23.0 years, SEM = 1.13) underwent all experiments in the fMRI scanner. All participants had normal or corrected-to-normal vision, were right handed, and reported no history of psychiatric or neurological disorders or family history of such conditions. None had used psychotropic medications in the past 6 months. The sample size followed previous VWM decoding studies that have demonstrated reliable effects with cohorts of comparable size (3, 21). Crucially, an extensive amount of fMRI data was collected from each participant, consisting of six different tasks across more than 12 scanning sessions, containing 1520 event-related VWM trials. This sampling design maximizes statistical power for within-subject comparisons and allows us to probe general object representation properties while maintaining identical stimulus sets across participants. The resulting rich data per participant provide a rigorous foundation for task-level contrasts and support the generalizability of our conclusions despite the modest cohort size. This study was approved by the Ethics Committee of School of Psychological and Cognitive Sciences at Peking University (project ID: 2022-03-04). Written informed consent was obtained from each participant before the experiment, and they were compensated for their participation.

Stimuli and experimental procedure

All six participants completed six experimental tasks, including three perception tasks, two VWM tasks, and one attention task. Typically, each participant first completed two perception tasks (CP task and BP task), followed by two VWM tasks (one-item WM and two-item WM tasks), and then the UP task and the BA task, across 12 to 16 separate scanning sessions. Each scanning session lasted 0.75 to 1.5 hours. In addition, two localizer tasks were performed before the experimental tasks. All experiments were generated via Psychtoolbox-3 (91–93) on MATLAB 2016b (MathWorks).

Across all tasks, 20 grayscale images of isolated real-world objects were used as stimuli (fig. S1A); face images were taken from the FEI Face Database (https://fei.edu.br/~cet/facedatabase.html). In the perception and attention tasks, stimuli were presented at a visual angle of 10° using a fast-event design. In the VWM tasks, stimuli appeared at 10° visual angle in the encoding array and 5° visual angle in the probe array using an event-related design. All stimuli were presented on a white background. Two display systems were used across participants and tasks: one back projected onto a translucent screen inside the scanner and the other displayed on a liquid-crystal display (LCD) screen outside the scanner bore. Both systems operated at a refresh rate of 60 Hz and a spatial resolution of 1024 × 768 pixels. Stimulus presentation programs were adjusted to ensure consistent visual angle across both display systems. During scanning, participants lay supine and viewed the stimuli via a mirror mounted on the head coil.

UP task

UP task (Fig. 1A): In each trial, a single object was presented on one side of the screen, with a 0.3° offset between the edge of the image and the central fixation. The object was flashed three times within 1 s (200 ms on, 200 ms off), followed by a 3-s blank. Each object was presented 20 times per side, resulting in 800 valid trials across 10 scanning runs. Each run consisted of 100 trials, including 80 valid trials and 20 empty trials, each lasting 4 s. The lead-in and lead-out times were both set to 6 s, resulting in total run duration of 412 s per run. Within each scanning run, participants were required to perform a fixation task, with the central fixation point alternating between “+” and “×” every 3 to 7 s. Participants were instructed to respond with a key press upon detecting a fixation change.

BP task

BP task (Fig. 1B): The presentation settings and task requirements were similar to the UP task, except that two objects were presented simultaneously. Each object appeared on each side of the screen, with a 0.3° gap between each object’s edge and the central fixation. All possible pairings of the 20 objects yielded 400 valid trials across four scanning runs. Each run contained 100 valid trials and 25 empty trials, each lasting 4 s, resulting in a total run duration of 512 s. Trials in which the same object appeared in both visual fields (n = 20) were excluded from analysis.

CP task

CP task (fig. S4A): The presentation settings were similar to the UP task, except that the objects were presented centrally. In each trial, the object was flashed three times, and in 10% of the trials, the contrast of the non-white pixels in the object was reduced by 50% during the middle flash. Instead of using a central fixation task, participants were instructed to press a key upon detecting the contrast change in the object. Each object was presented 15 times, yielding 300 valid trials across three runs. Each run contained 100 valid trials and 25 empty trials, each lasting 4 s, resulting in total run duration of 512 s.

One-item WM task

As shown in Fig. 2A, the one-item WM task was a modified single-item postcue working memory task based on previous studies (3, 20), incorporating a spatial postcue rather than a sequential one. In each trial, participants viewed the encoding array consisting of two objects simultaneously presented in the left and right visual fields for 0.2 s, with a 0.3° offset between the edge of the image and the center fixation (same as in the BP task). This was followed by a 2.25° visual angle arrow cue presented at the center indicating the object to be memorized. After a 10-s delay, a probe array containing four objects was presented, with the images offset by 2.5° from the horizontal and vertical midlines. Each position of the probe objects corresponded to a button on the response box, and participants were instructed to select the corresponding object via button press or withhold response if the cued object was absent. Feedback regarding accuracy of the current response was provided by a fixation color change following the button press.

The objects in the encoding array were composed of two different items selected from the 20 objects. To counterbalance the number of presentations of different objects, spatial locations, and cueing positions, all possible permutations were generated, ensuring that each of the 20 objects appeared an equal number of times as the cued or uncued object in both visual fields. This yielded a total of 760 trials per participant, with 19 trials for each object per condition

$$\text{All} \text{trial number}=\frac{20!}{(20-2)!}\times 2$$ (1)

As for the probe array, in 75% of the trials, the cued object appeared at one of the four possible screen positions, each with approximately equal probability (142 to 144 trials per position) to prevent response biases. In the remaining 25% of the trials, the cued object was absent. The probe array could contain the cued object, the uncued object, both, or neither from the sample array. Specifically, in about 56% of trials, both the cued and uncued objects appeared; in about 38% of trials, only one of them appeared (with equal probability for cued or uncued); and in about 6% of trials, neither appeared. The remaining objects in each probe array were randomly selected from a subset of 80 objects, which were manually selected from the same larger image set as the 20 objects (94). Thus, participants could not rely solely on long-term familiarity with the 20 objects to achieve high behavioral performance. All four objects in the probe array were unique for a certain trial. Each run consisted of 20 trials, each lasting 16 s, with intertrial intervals of 2, 4, 6, 8, or 10 s randomly distributed across trials. A 6-s lead-in period was included at the beginning of each run, with no lead-out time, resulting in total run duration of 446 s. Participants completed 38 scanning runs across five to six scanning sessions.

Two-item WM task

As shown in Fig. 2B, the two-item WM task was modified from the one-item WM task described above, differing only in the appearing time of the cue. We moved the cue to the end of the delay period, requiring participants to retain both objects throughout the delay. The trial settings were similar to the previous one-item WM task, with all object orders and combinations newly generated. This task also comprised 38 runs, each lasting 446 s, distributed across four to six scanning sessions for each participant to finish.

BA task

BA task (Fig. 4A). The attention task was adapted from the BP task, with the following modifications to increase attentional demands. Instead of a changing fixation point, a constant red dot was presented at the center. In each trial, objects also flashed three times, and during the middle frame, one object on either side rotated slightly. The rotation angle was controlled by the QUEST procedure (50) to converge on 70% accuracy. Participants needed to press one key for a leftward rotation and another key for a rightward rotation. Each run contained 100 valid trials and 25 empty trials, each lasting 4 s, resulting in total run duration of 512 s. Data from 20 valid trials where the same object appeared in both visual fields were discarded from analysis.

Localizer tasks

Retinotopic mapping

To map EVC regions for each participant, we conducted the pRF experiment and estimated pRF as described by previous studies (14, 95). Throughout the experiment, participants were required to perform a fixation color judgment task, pressing a button whenever they detected a change in color. Stimuli consisted of wedge, ring, and bar apertures composed of objects scattered on a pink-noise background. In the first two runs, wedges rotated clockwise and counterclockwise around the fixation point; in the next two runs, rings expanded outward or contracted inward from the fixation point; and in the final two runs, bars moved along the diameter with different directions. Stimuli subtended 16° of visual angle. Each pRF run lasted 300 s, with the aperture and images updating at 15 Hz.

High-level visual area localization

To define high-level visual regions, we used a category-selective localizer with images of bodies, faces, scenes, objects, words, and Fourier phase–scrambled images. The localizer task contained two block-design runs, each lasting 396 s and containing 24 blocks (6 categories × 4 repeats). The lead-in and lead-out times were both set to 6 s. Each block lasted 16 s. Within each block, 16 images from the same category were presented for 500 ms, followed by a 500-ms blank interval. Isolated objects or scenes were overlaid on a scrambled background and displayed centrally at a visual angle of 8°. Participants performed a one-back task, pressing a button when the current image matched the previous one to stay engaged with the task.

fMRI data acquisition and preprocessing

All fMRI data were collected using a 3T Siemens Prisma MRI scanner at the Center for MRI Research, Peking University, equipped with a 64-channel phased-array head coil. In each scanning session, a high-resolution T1-weighted anatomical image was acquired using a three-dimensional magnetization-prepared rapid gradient-echo sequence [TR = 2530 ms, echo time (TE) = 2.98 ms, inversion time = 1100 ms, flip angle = 7°, voxel size = 0.5 mm by 0.5 mm by 1.0 mm, parallel acquisition technique factor = 2, generalized autocalibrating partially parallel acquisitions (GRAPPA) reconstruction]. Functional images for all tasks were acquired using a T2*-weighted gradient-echo echo-planar imaging (EPI) sequence with the following parameters: TR = 1000 ms, TE = 30 ms, flip angle = 78°, voxel size = 2.5 mm by 2.5 mm by 2.5 mm, total slices = 56, multiband (simultaneous multi-slice) acceleration factor = 4, and GRAPPA phase-encoding acceleration = 1. Field maps were acquired before each functional run for distortion correction.

The data were preprocessed using the BrainVoyager QX software (version 2.8; Brain Innovations, Maastricht, The Netherlands) and custom MATLAB code in conjunction with the NeuroElf toolbox (https://neuroelf.net/). The anatomical image from the first scanning session of each participant was manually aligned to the anterior commissure-posterior commissure line and transformed into the Talairach space. All subsequent T1 images were aligned to the first Talairach-transformed structural image. Cortical surfaces were reconstructed by segmenting gray and white matter, followed by inflation of the white and gray matter boundary.

Functional images were initially corrected for EPI distortions using the field maps via the anatabacus plugin in BrainVoyager QX. Subsequent preprocessing included temporal correction, motion correction (trilinear/sinc interpolation), spatial smoothing (full width at half maximum = 4 mm), and high-pass temporal filtering [general linear model (GLM)–Fourier with two sine/cosine pairs]. The functional images closest in time to the T1 scan were used as a reference and aligned to the T1 image, and all other functional images were aligned to this reference. The first six TRs of each run were discarded to minimize the transient magnetic saturation effect. Next, BOLD signals were extracted from all the remaining data for each scanning run. For each vertex, the signals were normalized across all remaining TRs.

To project the results on a common surface space in the searchlight RSA, we used fMRIPrep 22.0.1 (96) to repreprocess all the data. Structural MRI images underwent intensity correction and skull stripping and were then normalized to the MNI152NLin2009cAsym template. Brain tissue segmentation into cerebrospinal fluid, white matter, and gray matter was performed using FSL, and brain surfaces were reconstructed with FreeSurfer. For functional MRI, each BOLD run was motion corrected, slice-time corrected, and then coregistered to the T1w reference. The data were then resampled to both standard and surface spaces using ANTs and FreeSurfer. Internal operations leveraged Nilearn for functional processing. All preprocessed data were transformed to BrainVoyager files via fp2bv toolbox (https://github.com/UW-CHN/fp2bv).

For the perception and attention tasks, a GLM was used to estimate the beta responses for each trial via custom MATLAB scripts. For the VWM tasks and retinotopic tasks, raw BOLD responses were used for analysis. Data in each run were normalized across TRs for each vertex. For the block-design localizer task, we used voxel-wise GLMs within BrainVoyager QX software to calculate the contrast maps, and these volumetric maps were subsequently projected onto the individual cortical surface.

ROI definitions

All ROIs were defined in the participant-specific surface space using functional localizers or established atlases. For high-level visual areas, the OSC was defined using data from the localizer task as described above. A contrast of (face + body + object + scene – 4 × scrambled) was used to define the object-selective regions. Significantly activated vertices [q(FDR) < 0.05] near the lateral occipital and ventral temporal cortices were selected (30, 97), guided by the lateral occipital cortex (inferior division) and fusiform gyrus parcels from the Harvard-Oxford cortical atlas of the FSL package (98). Individual participant OSC ROIs are shown in fig. S18.

We used data from retinotopic mapping to localize early visual areas. The pRF estimation was calculated using scripts provided by Benson et al. (14) and Kay et al. (95). On the basis of the pRF results, we defined V1 to V3 as in Benson et al. (14). These regions were then combined together to form the EVC ROI.

For association areas, the IPS and FEF were defined using the probabilistic atlas of visual topography from Wang et al. (99), which is based on retinotopic mapping data from a large independent sample. Probability maps for IPS0 to IPS5 and FEF, originally in Montreal Neurological Institute (MNI) space, were transformed to each participant’s anatomical space and projected onto their cortical surface. Vertices with a probability greater than 10% were selected, and all IPS subregions (IPS0 to IPS5) were combined to form the IPS ROI. The FEF ROI was defined using the same 10% probability threshold. Critically, these atlas-derived ROIs are fully independent of the task data used for subsequent analyses.

Multivariate pattern analysis

We used decoding and RSA to investigate the neural representational patterns. In VWM tasks, BOLD time series data were extracted from each ROI and segmented into 18-s segments based on trial onset. In the perceptual and attention tasks, beta values from each trial were used. All trials within each condition were used as samples, with vertex-wise BOLD or beta values within each ROI serving as features for subsequent analysis.

Decoding analysis

In the main decoding analysis, a one-versus-rest multiclass classifier (20 classes) was constructed using a linear SVM algorithm based on Python’s scikit-learn library. Leave-one-trial-out cross-validation was used, training the model on n − 1 samples and testing on the held-out sample. Classification accuracies from the n folds were averaged to obtain the final decoding accuracy (chance level = 0.05). In addition, for each task condition, pairwise binary decoding was performed for all object pairs. The resulting decoding accuracies were used to construct a decoding-based RDM. The averaged decoding accuracy across all objects was calculated using the averaged off-diagonal values in the RDM (chance level = 0.5).

In all displayed results, decoding accuracies were shown separately for the contralateral and ipsilateral conditions. Specifically, in perception and attention tasks, the decoding accuracy for the contralateral and ipsilateral conditions was calculated as

$${\text{Acc}}_{\text{contralateral}}=\frac{1}{2}({\text{Acc}}_{\text{LHRV}}+{\text{Acc}}_{\text{RHLV}})$$

$${\text{Acc}}_{\text{ipsilateral}}=\frac{1}{2}({\text{Acc}}_{\text{LHLV}}+{\text{Acc}}_{\text{RHRV}})$$ (2)

where LHRV refers to decoding object labels in the right visual field (RV) using beta values from the left hemisphere (LH) ROI and RHLV refers to decoding object labels in the left visual field (LV) using beta values of each trial from the right hemisphere (RH) ROI. Similarly, LHLV and RHRV correspond to decoding object labels within the same-side hemisphere and visual field (i.e., ipsilateral conditions).

For the one-item postcue WM task, the full set of 760 trials was split into two sets, each with 380 trials based on cue direction (left or right). Then, the decoding accuracy for each condition was calculated as

$${\text{Acc}}_{\text{contralateral},\text{cued}}=\frac{1}{2}({\text{Acc}}_{\text{LHRV},\text{cue} \text{to right}}+{\text{Acc}}_{\text{RHLV},\text{cue} \text{to left}})$$

$${\text{Acc}}_{\text{contralateral},\text{uncued}}=\frac{1}{2}({\text{Acc}}_{\text{LHRV},\text{cue} \text{to left}}+{\text{Acc}}_{\text{RHLV},\text{cue} \text{to right}})$$

$${\text{Acc}}_{\text{ipsilateral},\text{cued}}=\frac{1}{2}({\text{Acc}}_{\text{LHLV},\text{cue} \text{to left}}+{\text{Acc}}_{\text{RHRV},\text{cue} \text{to right}})$$

$${\text{Acc}}_{\text{ipsilateral},\text{uncued}}=\frac{1}{2}({\text{Acc}}_{\text{LHLV},\text{cue} \text{to right}}+{\text{Acc}}_{\text{RHRV},\text{cue} \text{to left}})$$ (3)

where the LHRV and RHLV conditions represent decoding object identity in the right and left visual fields using LH and RH ROIs, respectively. Each subcondition (e.g., LHRV and cue right) included 380 of 760 trials with the cue directed to the relevant side. For example, under the contralateral, cued condition, the LHRV subcondition used BOLD signals from the LH ROI to decode object identity in the right visual field during trials cued to the right. Similarly, the RHLV subcondition used the other 380 trials cued to the left, with RH ROI BOLD signals decoding the object label in the left visual field. The decoding accuracy for the contralateral, cued condition was computed as the average of these two subconditions.

For the two-item postcue WM task, no cue information was available during the delay period; thus, both objects had equal mnemonic status. All 760 trials were used together to calculate decoding accuracy. Decoding accuracy was computed in the same manner as in the perception and attention tasks, estimating contralateral and ipsilateral decoding using Eq. 2.

Cross-lateral decoding analysis

To address the potential confound that decoding differences could be driven by variations in SNR rather than changes in representation (45), we performed an additional cross-lateral decoding analysis. In this analysis, the classifier was always trained on contralateral conditions and tested on ipsilateral conditions. All classifier settings were identical to those described in the above decoding analysis.

For each ROI and hemisphere, a linear SVM was trained on trials where the decoded object appeared in the contralateral visual field (e.g., LHRV for the LH and RHLV for the RH) and tested on trials where the object appeared in the ipsilateral visual field (e.g., LHLV and RHRV, respectively).

Specifically, in the UP task, half of the trials presented the object in the left visual field and the other half in the right visual field. For each hemisphere, the classifier was trained on contralateral trials and tested on ipsilateral trials. The cross-lateral decoding accuracy was calculated as

$${\text{Acc}}_{\text{cross}}=\frac{1}{2}({\text{Acc}}_{\text{train} \text{on} \text{LHRV},\text{test} \text{on} \text{LHLV}}+{\text{Acc}}_{\text{train} \text{on} \text{RHLV},\text{test} \text{on} \text{RHRV}})$$ (4)

For tasks with bilateral stimulus presentation (BP, BA, and the two-item postcue WM task), both visual fields contained stimuli in each trial. For each hemisphere, the contralateral object in each trial was used as the decoding label. Cross-lateral decoding was performed using leave-one-trial-out cross-validation: On each fold, one trial was held out for testing, the classifier was trained on the contralateral objects from the remaining trials, and generalization was tested on the ipsilateral object from the held-out trial. Cross-lateral decoding accuracy was computed as in Eq. 4.

For the one-item postcue WM task, trials were separated by cueing condition. The decoder was trained on trials in which the cued item appeared in the contralateral visual field and tested on trials in which the cued item appeared in the ipsilateral visual field. Cross-lateral decoding accuracy was computed separately for cued and uncued conditions

$$\begin{array}{c}{\text{Acc}}_{\text{cross},\text{cued}}=\hfill \\ \frac{1}{2}({\text{Acc}}_{\text{train} \text{on} \text{LHRV} \text{cue} \text{to right}, \text{test} \text{on} \text{LHLV} \text{cue} \text{to left}}+{\text{Acc}}_{\text{train} \text{on} \text{RHLV} \text{cue} \text{to left}, \text{test} \text{on} \text{RHRV} \text{cue} \text{to right}})\hfill \end{array}$$

$$\begin{array}{c}{\text{Acc}}_{\text{cross},\text{uncued}}=\hfill \\ \frac{1}{2}({\text{Acc}}_{\text{train} \text{on} \text{LHRV} \text{cue} \text{to left}, \text{test} \text{on} \text{LHLV} \text{cue} \text{to right}}+{\text{Acc}}_{\text{train} \text{on} \text{RHLV} \text{cue} \text{to right}, \text{test} \text{on} \text{RHRV} \text{cue} \text{to left}})\hfill \end{array}$$ (5)

Vertex-ablation analysis

To assess the proportion of OSC vertices contributing to object identity representation, we performed a vertex-ablation analysis using beta values from the UP task. The goal was to isolate and exclude vertices with similar response profiles across the two visual fields, which may reflect inherently bilateral visual input. For each vertex in the OSC, the correlation between its beta values for the 20 objects under the ipsilateral and contralateral conditions was calculated and served as the bilateral responsivity

$$\text{Bilateral responsivity}=\text{Corr}({\text{beta}}_{\text{LHLV}},{\text{beta}}_{\text{LHRV}})$$ (6)

We performed stepwise vertex-ablation by progressively excluding the vertices with the highest bilateral responsivity for both the contralateral and ipsilateral condition. At each step, an increasing percentage of vertices (N%) were excluded, and decoding accuracy was recalculated on the basis of the remaining vertices.

Representational similarity analysis

To test for the representational pattern similarities across tasks, we conducted RSA based on the task-specific RDMs. For each condition, we first constructed the RDM by averaging the beta values across all trials for each object label and then computed the pairwise Pearson correlation R(object A, object B) between every object pair using the beta values from the ROI vertices. The dissimilarity between object pairs was computed as 1 – R(object A, object B), resulting in a 20 × 20 symmetric RDM for each condition. For the CP task, RDMs were calculated by combining beta values from ROIs in both hemispheres; for all other tasks, RDMs were calculated using beta values or BOLD signals from a single-hemisphere ROI. RDMs under different conditions were calculated on the basis of the data and labels described above. Representational similarity was quantified as the Pearson correlation between the off-diagonal values of two corresponding RDMs. Typically, each condition’s RDM was compared against the CP task RDM, which served as a common sensory template. Contralateral and ipsilateral RSA results were obtained by averaging these correlations across subconditions (Eq. 2–3). All results were calculated at the individual level and Fisher Z transformed before group-level statistical comparisons.

Searchlight-based RSA

Searchlight RSA was performed in both participants’ native surface space and the fsaverage space. We first constructed reference RDMs and floating RDMs. The reference RDMs were derived from the OSC in the CP task (fig. S4), where objects were presented at the center. Each floating RDM was calculated for each vertex using the BOLD or beta values from the 200 surrounding neighbor vertices in the native surface space and 400 surrounding neighbor vertices in the fsaverage space. Neighbor vertices numbers of 100, 200, and 300 were also tested and did not produce systematic differences from the displayed results. In the two VWM tasks, we used the average BOLD signal from the 6 to 10 TRs for constructing VWM RDMs; in all other tasks, trial-wise beta values were used. To assess representational similarity for each vertex, we computed the Pearson correlation between the off-diagonal values of the reference RDM and each floating RDM. In Fig. 5, we present the representational similarity for ipsilateral condition, comparing the bilateral OSC reference RDM to the floating RDMs under the ipsilateral conditions. The calculation followed the same procedure used in the decoding and RSA analyses. Specifically, for the LH, the RSA results are derived from the LHLV subcondition (LH ROI, left visual field), and for the RH, the RSA results are from the RHRV subcondition (RH ROI, right visual field). The RSA was calculated at individual level using their individual reference RDM and floating RDMs for each task.

For visualization on the common surface space, the representational similarity values were averaged across all participants in the fsaverage space. For statistical analysis (fig. S15), r values were extracted from the OSC for each participant in native surface space, and all values were Fisher Z transformed. All Fisher Z-transformed correlation values were pooled across participants to form the population distribution. In addition, participant-level averages were computed and used in statistical analyses.

Cross-ROI partial RSA

The cross-ROI partial RSA was performed using decoding-based RDMs calculated as described above. For the VWM tasks, we computed partial correlations between the ipsilateral and contralateral OSC RDMs and all other ipsilateral or contralateral ROI RDM, with the remaining RDMs serving as covariates. Pearson correlation coefficients were used and subsequently Fisher Z transformed for statistical testing.

Statistical analysis

All analyses used within-subject statistical tests. We reported parametric statistics (one-sample and paired t tests, and repeated-measures ANOVAs for interaction effects) in the Results, with parallel nonparametric Wilcoxon signed-rank test results, effect sizes (Cohen’s d), Bayes factors (BF₁₀), and estimated statistical power in the Supplementary Materials using Pingouin (100). In the decoding analysis, we tested whether decoding accuracy under each condition exceeded chance level. For the ablation decoding, the same statistical tests were applied for each ablation step. For the VWM cueing effect, we tested the interaction between laterality and cueing using a repeated-measures ANOVA; significant interactions were followed by simple-effect comparisons to assess whether decoding was higher for cued versus uncued objects and whether each condition exceeded chance. For RSA, Fisher Z-transformed values were tested against zero. To directly compare representational strength across tasks, we performed paired t tests on decoding accuracies and Fisher Z-transformed RSA correlations. To quantify regional dissociations in representational format, we conducted a two-way repeated-measures ANOVA (region × laterality) on the Fisher Z-transformed RSA correlations, with significant interactions followed by simple-effect analyses. Unless otherwise noted, tests were one tailed.

To control for the false-positive rate, we applied Benjamini-Hochberg FDR correction. Specifically, FDR correction was applied to the following: (i) the two one-sample decoding tests (contralateral versus chance and ipsilateral versus chance) within each task and ROI; (ii) the three simple-effect tests within each hemisphere for the cueing analysis (cued versus uncued, cued versus chance, uncued versus chance); (iii) the 12 ablation-step tests for each task in the vertex-ablation analysis; (iv) the seven between-task contrasts evaluating ipsilateral representation; and (v) the relevant sets of one-sample RSA tests and one-sample decoding tests when multiple conditions were considered together (e.g., figs. S2, S13, and S15 to S17). Original P values are reported in the Supplementary Materials.

Supplementary Materials

This PDF file includes:

Figs. S1 to S18

Tables S1 to S12