(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



Collaborative rule learning promotes interbrain information alignment [1]

['Denise Moerel', 'The Marcs Institute For Brain', 'Behaviour', 'Development', 'Western Sydney University', 'Sydney', 'Tijl Grootswagers', 'School Of Computer', 'Data', 'Mathematical Sciences']

Date: 2025-12 

Social interactions shape our perception of the world, influencing how we interpret incoming information. Alignment between interacting individuals’ sensory and cognitive processes is key to successful cooperation and communication, but the neural processes underlying this alignment remain unknown. Here, we leveraged Representational Similarity Analysis (RSA) on electroencephalography (EEG) hyperscanning data to investigate information alignment in 24 pairs of participants who performed a categorization task together based on agreed-upon rules. Significant interbrain information alignment emerged within 45 ms of stimulus presentation and persisted for hundreds of milliseconds. Early alignment (45–180 ms) occurred in both real and randomly matched pseudo-pairs, reflecting shared sensory responses. Importantly, alignment after 200 ms strengthened with practice and was unique to real pairs, driven by shared representations associated with, and extending beyond, the categorization rules they formed. Together, these findings highlight distinct processes underpinning interbrain information alignment during social interactions, that can be effectively captured and disentangled with Interbrain RSA.

(A) Participants were seated back-to-back, facing separate but identical displays. Each participant had a keyboard to provide responses. (B) There were 16 unique grayscale stimuli, varying on 4 dimensions: spatial frequency (thick vs. thin lines), pattern (wavy vs. straight lines), contrast (high vs. low), and general shape (circle vs. square). (C) The two-part categorization task. In Part 1 (top panel), participants worked together to sort the stimuli into 4 groups of 4 stimuli. To do this, they had to select two dimensions to group the stimuli, while ignoring the two other dimensions. In Part 2 (bottom panel), participants categorized stimuli into the four groups they just created: Participants saw a stimulus for 100 ms, followed by a blank screen for 900 ms. Then, a response-button mapping screen was shown for 1,000 ms, and participants could make a categorization response based on the four groups that were made in Part 1. This was followed by a feedback screen for 500 ms, providing feedback about whether the response of the two participants matched the previously agreed-upon sorting rules. The feedback pertained to the pair, not the individual. During the feedback, the fixation bullseye turned green if both individuals in the pair were correct, red if neither individual was correct, and orange if only one individual was correct. In the latter case, the feedback was not specific about which individual in the pair was correct. (D) The features that were used by the pairs to make the four groups. The x-axis shows the percentage of pairs that chose each feature. (E) The combinations of features that were used to make the four groups. Each pair of participants chose two features. The x-axis shows the percentage of pairs that chose each combination of features. (F) Behavioral accuracy on the categorization task over experiment blocks. The thick blue line shows the average across the 48 individuals. The shaded blue area shows the 95% confidence interval. The thin gray lines show the accuracy of individual participants. Chance is at 25% in this 4-way categorization task (dotted line). ( G) Response alignment between the two participants in the pair, regardless of whether the response was correct. The green line shows the average alignment across the 24 pairs. The shaded green area shows the 95% confidence interval, and the dotted line indicates chance level (25%).

In this study, we used Interbrain RSA with EEG hyperscanning to investigate the emergence and dynamics of information alignment between interacting individuals’ brains while presented with various visual stimuli. We aimed to distinguish information alignment that was purely evoked from participants seeing the same thing at the same time, from alignment that was driven by each pair’s pre-formed rules and interpretations. We recorded EEG data from 24 pairs of participants sitting back-to-back while performing a 4-way categorization task on visual stimuli. Importantly, the two participants within each pair had previously created the four groups together by agreeing on arbitrary categorization rules (see Fig 1 ). We used Interbrain RSA to determine the time-course of neural information alignment between individuals within a pair during the categorization task. To determine whether and when the neural information alignment was purely evoked by shared visual information (seeing the same stimulus), we compared Interbrain RSA between real participant pairs and randomly matched-up pseudo-pairs who saw the same stimuli in the same order. Finally, we used randomly matched-up pseudo-pairs who independently created the same four groups of stimuli to determine the extent to which the alignment was unique to the pair or driven by the pair’s pre-formed rules. We expected alignment, beyond that driven by shared visual input, to be reflected by higher Interbrain RSA in real pairs, which would occur during later information processing stages, be more widely distributed across the brains, and strengthen over time with practice and integration of pre-agreed categorization rules. Moreover, we examined whether this neural information alignment persisted beyond the categorization task, which we tested by recording EEG data during passive viewing of the same stimuli in a rapid visual stream before and after participants agreed on the categorization rules and performed the categorization task.

Interbrain RSA holds promise to capture and understand how visual representations evolve and converge across individuals through social interactions. However, its capacity to uncover information alignment beyond that simply driven by robust differences in evoked sensory responses, as modulated by same versus different visual objects in Varlet and Grootswagers [ 13 ], remains to be demonstrated. Indeed, understanding how the same visual stimuli are (dis)similarly represented by individuals depending on the groups they belong to critically requires not only being able to track information alignment that is reliably and externally evoked, but also alignment that is socially learnt and internally induced [ 27 ]. The latter type of information alignment is expected to reflect convergence in higher-level representations, extending beyond initial visual responses occurring as early as ~50 ms after stimulus onset [ 31 – 33 ], and engaging more distributed brain networks involved in attention and decision-making. Consistent with this notion, previous studies have demonstrated task- and decision-relevant modulations of visual information processing from 140 to 220 ms onwards [ 34 – 38 ] and within frontal-parietal networks [ 37 , 39 – 41 ]. Regions across the frontal cortex, including the anterior cingulate and medial prefrontal cortices, have also been implicated in hyperscanning research during real-time social interactions [ 22 ].

One study found evidence for coincidental synchrony between pseudo-pairs using Phase-Locking Value and to a lesser degree Coherence and Partial Directed Coherence measures [ 25 ]. This coincidental synchrony has been attributed to common EEG rhythmicity across participants, arising from systematic differences between experimental conditions. Another study provided evidence that methodological decisions, such as the use of short epochs, can inflate phase-based estimates of interbrain synchrony [ 29 ]. Finally, recent work highlighted that phase- as well as amplitude-based measures of interbrain synchrony were not sensitive to the alignment of information across people, as they were minimally affected by whether participants saw the same or different objects [ 13 ]. The authors introduced Interbrain Representational Similarity Analysis (RSA), an adaptation of RSA [ 30 ], to more effectively capture and compare the information content represented in each brain. Unlike synchrony measures, Interbrain RSA is information-based, allowing researchers to identify what information is shared between individuals at specific moments in time. It also provides a single, experiment-level estimate of information alignment between brains, making it less susceptible to transient fluctuations driven by shared environmental factors, such as changes in lighting or acoustic conditions. Finally, by abstracting away from the raw data and focusing on the representational structure contained in the data, Interbrain RSA enables comparisons across different brain regions, time scales, neuroimaging modalities, and tasks.

Previous hyperscanning research has revealed enhanced synchrony between the neural oscillations of individuals as they engage in joint activities, which has been argued to support communication, coordination, and collaboration (for an overview, see [ 9 , 15 – 17 ]. Higher neural alignment between people, as indexed by a range of measures of synchrony (e.g., Coherence, Phase Locking Value, weighted Phase Lag Index), has been shown during genuine social interactions—participants interacting together in real-time versus participants being paired a posteriori (pseudo-pairs; [ 18 , 19 ])—and with higher levels of familiarity (e.g., mother–infant interaction; [ 20 , 21 ]) and collaborations (e.g., collaborative versus competitive scenarios; [ 22 – 24 ]). However, several recent studies have questioned the nature of the processes captured by interbrain synchrony measures [ 25 – 28 ], and revealed critical methodological limitations [ 13 , 25 , 29 ].

The groups we belong to can significantly shape our perception and interpretation of the world around us. Whether these groups are based on culture, profession, social circles, or any other common interest, they provide a shared context, set of values, and ways of interpreting incoming information, which can facilitate communication and joint activities across our society [ 1 – 3 ]. Even though visual perception is largely automatic and seemingly objective [ 4 – 6 ], the interpretation of visual information can be strongly influenced by social interactions, modulating how similarly people represent and respond to their environment depending on whether they belong to same or different groups [ 1 , 7 , 8 ]. Hyperscanning, a technique that allows for the simultaneous measurement of brain activity in multiple people while they interact with each other, has become a valuable tool for studying real-time social interactions and the dynamics of interacting brains [ 9 – 11 ]. While previous hyperscanning research has focused on interbrain synchrony measures, recent evidence suggests that these measures are content agnostic and do not capture information alignment across people, and that multivariate representational analyses are a powerful alternative [ 12 – 14 ]. Here, we leverage and advance electroencephalography (EEG) hyperscanning methods to uncover how representations in interacting individuals’ brains evolve in real-time and uniquely converge as a function of commonly agreed upon-rules and interpretation of their visual environment.

2. Results

2.1. Commonly agreed-upon rules and behavioral alignment We recorded the brain activity of 24 pairs of participants using the 64-channel BioSemi active electrode system (BioSemi, Amsterdam, The Netherlands). The two participants in the pair were seated back-to-back in the same dimly lit room, viewing separate computer screens that showed identical visual input throughout the task (Fig 1A). Participants engaged in a joint categorization task, which consisted of two parts. In the first part, the two participants in the pair agreed upon sorting rules for visual stimuli, and in the second part, participants used these shared rules to categorize the images. There were 16 images that varied across 4 dimensions, with 2 levels per dimension (Fig 1B): spatial frequency (thick vs. thin lines), pattern (wavy vs. straight lines), contrast (high vs. low), and general shape (circle versus square). During the first part of the task, the two participants in the pair worked together to sort the images into 4 groups of 4, effectively selecting two of the stimulus dimensions while ignoring the other two dimensions (Fig 1C, top panel). Participants were allowed to talk during this stage. They saw identical screens and shared a cursor, which means they could both drag and drop the stimuli and observe the movements of the other participant. Both participants had to agree on the arbitrary rules before continuing the task. To determine whether the different pairs came up with different strategies, we analyzed which rules were used by each pair (Fig 1D). All dimensions were used across the 24 pairs: spatial frequency was chosen most often (58.33% of pairs), followed by contrast (54.17% of pairs), pattern (45.83% of pairs), and shape (41.67% of pairs). We also assessed which combination of two dimensions was used by the pairs (Fig 1E). Again, all combinations of dimensions were used, with the largest group choosing spatial frequency and contrast (33.33% of pairs), followed by spatial frequency and pattern as well as pattern and shape (20.83% of pairs), contrast and shape (16.67% of pairs), and by spatial frequency and shape as well as pattern and contrast (4.16% of pairs). These results show that all dimensions were used across pairs as classification rules, with a reasonable spread across all options. This means that the stimulus sorting rules were indeed arbitrary, in that different pairs of individuals derived different sets of rules during the communication phase. During the second part of the categorization task, the participants categorized the stimuli based on the rules they had just agreed upon by allocating each stimulus to one of the four newly-defined groups (Fig 1C, bottom panel). Participants would see a stimulus for 100 ms, followed by a blank screen for 900 ms, during which participants could make a decision about which category the stimulus belonged to. After the blank screen, four button responses were shown for 1,000 ms, which both participants had to select using four different keys on their keyboard to provide a categorization response. The mapping between the four buttons and four categories was randomly changed on every trial. This means participants could not know which key to press before the response mapping was presented on the screen. This delayed response screen with unpredictable button mappings ensured that the 1,000 ms of EEG data after stimulus onset—our window of interest—were free from any motor preparation and contamination. Participants were asked not to talk during this part of the task, to avoid contamination of the EEG signal, but they were allowed to talk during the self-paced breaks. Breaks occurred after every 32 trials, with less than 3 minutes between two consecutive breaks, and there were a total of 19 breaks throughout the task. We calculated the behavioral accuracy at the level of the individual (Fig 1F), by determining the percentage of responses that matched with the previously agreed-upon sorting rules. In addition, we calculated the alignment in behavioral responses between the two individuals in the pair, by determining whether the two individuals in the pair provided the same response, regardless of whether it was correct or incorrect (Fig 1G). Visual inspection reveals an upward trend for both behavioral measures, with higher accuracy as well as higher response alignment later in the experiment.

2.2. Evidence for information alignment beyond shared visual input We used Interbrain RSA to assess the degree to which neural representations were aligned between the two individuals in the pair during the categorization task. For each participant separately, we created a Representational Dissimilarity Matrix (RDM) for each time-point in the epoch, reflecting the dissimilarity in the pattern of activation across all 64 channels for each combination of two unique stimuli. To determine the time-course of information alignment between the two individuals in the pair, accounting for information alignment at different time-points between individuals, we calculated the correlation between the dissimilarity matrices of the two individuals for each combination of time-points in the epoch. Fig 2 (left panel) shows the information alignment between the participants in a pair, averaged across pairs. To determine whether the observed information alignment was due to simply viewing the same visual input, we obtained the same measure for randomly matched pseudo-pairs, based on 10,000 random pair assignments. Importantly, we matched the counterbalancing of the trial orders for the pseudo-pairs, with 4 unique trial orders among the 24 pairs. This ensured that the individuals in the pseudo-pairs saw the same stimuli, in the same order, as the real pairs. Fig 2 (right panel) shows the information alignment for the pseudo-pairs. Visual inspection shows a remarkably similar pattern of strong information alignment early in time, from approximately 50–165 ms after stimulus onset. However, the information alignment appears more sustained over time for the real pairs compared to the randomly matched pseudo-pairs. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 2. Information alignment between the two individuals in each pair. The left panel shows the mean (Spearman) correlation between the neural representation of the two individuals in each real pair for all combinations of time-points. The brightness of the color reflects the strength of the correlation, with brighter colors reflecting higher correlations, and therefore stronger information alignment. The y-axis shows the time for person A in the pair, and the x-axis the time for person B in the pair. The right panel shows the information alignment for the pseudo-pairs, who did not perform the task together. Here, we repeated the same analysis for randomly matched pseudo-pairs instead of real pairs and with results showing the average across 10,000 iterations. https://doi.org/10.1371/journal.pbio.3003479.g002 To reduce dimensionality and obtain a single time-course of information alignment, we averaged over one time dimension of the symmetrical time-by-time matrix (Fig 2), enabling us to capture information alignment across all possible time lags between the two participants (Fig 3A). The time-course shows the alignment for a given time-point of person A with all other time-points for person B in the pair and vice versa. Cluster-based permutation testing revealed significant alignment in the real pairs from 45 to 285 ms and 315 to 425 ms after stimulus onset. This means that during this time, the information alignment between individuals in a pair was stronger than would be expected by chance. There was no significant difference between the real pairs and randomly matched pseudo-pairs, who saw the same stimuli in the same order, from 45 to 180 ms (no significant clusters and all p values >.05). This suggests that early in time, the information alignment between participants was driven by shared visual information. However, there was a significant difference between the real pairs and randomly matched pseudo-pairs from 185 to 240 ms, 325 to 360 ms, and 370 to 425 ms after stimulus onset, with stronger information alignment observed for the real pairs compared to pseudo-pairs. This suggests that later in time, there was information alignment between individuals in the real pairs that was not purely visually evoked but uniquely emerged from commonly agreed-upon rules and interpretations. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 3. The time-course of information alignment. (A) The time-course of information alignment is shown in dark blue. This measure was obtained by collapsing the temporal generalization matrix of dissimilarity matrix correlations between the two individuals in each pair (Fig 2) along one of the time dimensions. Cluster corrected p values <.05 for this correlation are shown in dark blue below. The shaded area shows the 95% confidence interval. The time-course of dissimilarity matrix correlations for the randomly matched pseudo-pairs is shown as a dashed light blue line. This shows the information alignment that is driven by sensory-evoked signals. Cluster corrected p values <.05 for the difference between the real and pseudo-pairs are shown in red below. This shows the information alignment that was not driven by sensory-evoked signals. The topographies for the information alignment are shown below the plot. They show the average correlation of one channel with all channels for the other participant, averaged across 100-ms time bins. The top row shows the topographies for the real pairs, the middle row for the pseudo-pairs, and the bottom row for the contrast (real—pseudo-pairs). (B) The channels included in the six clusters: Frontal Left (F L), Frontal Right (F R), Temporal/Central Left (TC L), Temporal/Central Right (TC R), Parietal/Occipital Left (PO L), and Parietal/Occipital Right (PO R). (C) The dissimilarity matrix correlations between person A (left) and person B (right) in the pair, for all combinations of six channel clusters. Correlations are averaged across an early (45–150 ms) and late (200–1,000 ms) time-window. The color of the lines reflects the strength of the correlation, with darker color reflecting stronger information alignment. The left plots show the correlation within real pairs, the middle plots show the correlation for the pseudo-pairs, and the right plots show the difference between real and pseudo-pairs. https://doi.org/10.1371/journal.pbio.3003479.g003 To determine which channels were driving the information alignment between individuals in a pair, we calculated for each participant the correlation at each channel with all channels from the other participant. The dissimilarity matrices for each channel, used to obtain the time-by-time map, were computed using 4 or 5 neighboring channels. The single time-course correlation obtained from these dissimilarity matrices was then averaged into 100-ms intervals to obtain scalp topographies. These topographies indicated for each channel and time bin of participant A, how strong the alignment was with any of the other channels at any time bin for participant B and vice versa (Fig 3A). They showed that before 100 ms, the alignment was mostly driven by posterior channels, moving to more frontal channels between 100 and 200 ms. Importantly, the alignment in real and pseudo-pairs looked very similar during this time. Around 300–400 ms, visual inspection appeared to show stronger alignment for real compared to pseudo-pairs. This latter alignment was driven by channels that were distributed across the scalp. To further explore the spatial dynamics of the information alignment, and identify the most aligned regions across brains, we repeated the same analysis used for the topographies, but this time keeping the combinations of channels for both brains and grouping them into six pre-defined clusters to reduce dimensionalities (Fig 3B). We collapsed the time-course into early time-points (45–150 ms) and late time-points (200–1000 ms). The resulting channel cluster correlations (Fig 3C) indicated for each cluster of channels and time bin of participant A, how strong the alignment was with individual clusters of channels for participant B in the same time bin. The results show that for the early time window, there was information alignment for all clusters of channels, with the strongest alignment between left and right parietal/occipital channels (Fig 3C). Visual inspection reveals a similar pattern for the real and pseudo-pairs, adding further evidence that early alignment is driven by visual information. For the late time window (200–1,000 ms), there was stronger information alignment for real compared to pseudo-pairs in channel clusters distributed across the scalp. The strongest alignment was observed between the left and right temporal/central clusters. These findings are in line with a distributed network of regions driving the later neural information alignment.

2.3. Late information alignment strengthens over the experiment The behavioral categorization data showed that participants became more accurate over the course of the experiment, with more similar responses given by the two individuals in the pairs in later blocks. We assessed whether this finding was mirrored in the information alignment between participants, by repeating the same Interbrain RSA analysis as described above separately for the first and second half of the experiment. Fig 4A shows the information alignment between participants in a pair for the two experiment halves. We found evidence for neural information alignment during the early time window for both experiment halves; between 65 and 185 ms after stimulus onset for the first experiment half and between 45 and 175 ms second experiment half. However, the information alignment in the later time window was more sustained for the second experiment half. For the first experiment half, we only found evidence for information alignment between 515 and 575 ms after stimulus onset, whereas for the second experiment half, we observed information alignment between 225 and 280 ms, and for much of the time-window between 400 and 900 ms. We calculated the difference between real and pseudo-pairs to assess the neural information alignment that was not driven by shared visual information (Fig 4B). The early information alignment disappeared, suggesting this alignment was simply evoked by visual information. The information alignment in the later time window was still present after subtracting the information alignment of the pseudo-pairs, suggesting that the later information alignment emerged from pair’s pre-formed sorting rules and common interpretation. Again, we found more sustained information alignment for the second experiment half. For the first experiment half, late information alignment was present between 510 and 575 ms and between 775 and 825 ms after stimulus onset. For the second experiment half, we found information alignment between 225 and 280, and for much of the time-window between 385 and 925 ms. Together, these results suggest that the early information alignment, which was evoked by visual information, did not change over the course of the experiment. However, later alignment, that was unique to real pairs and likely driven by higher-order cognitive processes such as attention and/or decision-making, was more persistent over time in later compared to earlier experiment trials. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 4. The time-course of information alignment, split by first and second experiment half. (A) The time-course of the information alignment between the two participants in each pair for the first experiment half (light blue), and the second experiment half (purple). The shaded areas around the plot lines show the 95% confidence intervals. Cluster corrected p values <.05 are shown below the plot in the same colors. (B) The difference in information alignment between the real pairs and pseudo-pairs (real—pseudo). This shows the alignment that is not driven by visual information. All plotting conventions are the same as Fig 4A. https://doi.org/10.1371/journal.pbio.3003479.g004

2.4. Beyond shared rules: Evidence for pair-specific information alignment To assess whether this late neural information alignment was driven by representations that were unique to the pair, beyond those driven by shared rules, we conducted an exploratory analysis to compare real pairs to pseudo-pairs who shared the same stimulus classification rules. This analysis was conducted separately for the first and second half of the experiment, and was based on 16 of the 24 pairs, as some pairs had no other pairs with the same trial order that used the same rules. Our results show that the late alignment was only partially driven by the dimensions participants used to categorize visual stimuli (Fig 5). Indeed, the late alignment in real pairs was significantly higher than that observed in pseudo-pairs where participants used the same dimensions to categorize the stimuli. Stronger alignment was already observed in the first half of the experiment, with a significant cluster from 335 to 435 ms, and this effect appeared to strengthen in the second half, with extended significant clusters between 570 and 875 ms. Interestingly, this stronger alignment in the second half was observed despite higher correlation also being observed for pseudo-pairs, likely reflecting the reinforcement of shared rules. These results suggest that late alignment was driven not only by shared categories collectively formed a priori but also by pair-specific representations emerging from real-time, trial-to-trial interactions through shared perceptions, decisions, collective feedback, and communication between blocks. Furthermore, these results indicate that the strengthening of this late alignment in the second half reflects both non-pair-specific (shared rules) and pair-specific information alignment. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 5. The contribution of shared rules to neural information alignment, for the first (A) and second (B) experiment half. All plotting conventions are the same for Fig 5A and 5B. The time-course of information alignment is shown in dark blue. This measure was obtained by collapsing the temporal generalization matrix of dissimilarity matrix correlations between the two individuals in each pair along one time dimension. The shaded area represents the 95% confidence interval. Significant correlations (cluster-corrected p <.05) are shown in dark blue below. The dashed orange line shows the time-course of correlations for same-rules pseudo-pairs, who are randomly matched individuals who independently came up with the same rules. This shows the information alignment that is driven by cognitive processes associated with the same rules, as well as sensory evoked signals. Significant differences between real pairs and same-rules pseudo-pairs (cluster-corrected p <.05) are shown in red below, highlighting the socially induced information alignment that cannot be explained by shared rules alone. Note this is based on 16 pairs (32 participants), who were also included in the same-rules pseudo-pairs. We excluded six pairs because there was no other pair within the same counterbalancing that picked the same categorization rules. https://doi.org/10.1371/journal.pbio.3003479.g005

[END]
---
[1] Url: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3003479

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/