Bidirectional information flow in cooperative learning reflects emergent leadership

Abstract

Advances in social neuroscience have shown that one of the fundamental characteristics of cooperative learning is synchronization between learners’ brains. However, the directionality of this synchronization, and the role of emergent leadership (i.e., a group leader emerges naturally), in cooperative learning remain unclear. Here, we investigated the directionality and dynamics of information flow by leveraging functional near-infrared spectroscopy (fNIRS) hyperscanning and Granger causality analysis (GCA). Through a 6 min dyadic cooperative learning task, we observed that dyads’ utterance score increased over time and remained stable at the end of interaction, suggesting successful cooperative learning. At the neural level, we found a stronger leader-to-follower Granger causality in the left middle temporal gyrus, alongside a more pronounced follower-to-leader causality in the left sensorimotor cortex. Moreover, we found that information transfer in both directions increased and peaked around the first half of time into the task, followed by a decline. These temporally similar yet spatially dissociable patterns of directional information flow suggest a hierarchical organization of bidirectional communication during cooperative learning with emergent leadership.

Using fNIRS hyperscanning and Granger Causality analysis, this study unveiled a hierarchical organization of bidirectional communication during cooperative learning with emergent leadership.

Introduction

Confucius, an ancient Chinese educator, once remarked, “Studying alone without peers leads to limited knowledge and narrow perspectives (独学而无友, 则孤陋而寡闻)”. This wisdom has evolved into a well-accepted pedagogical approach known as cooperative learning. This approach involves students engaging in learning activities within small groups, where rewards or recognition are based on the group’s performance¹. Over the past 70 years, the practice of cooperative learning has represented a shift towards an educational environment that values cooperation and mutual support, rather than a survival-of-the-fittest mentality and rugged individualism². While we have substantial understanding of how peer interactions shape personal growth, we know less about how these learner-to-learner interactions are reflected and organized in (and across) the brain(s).

Our recent observations align with previous studies suggesting that learners’ brain activity becomes more synchronized during cooperative learning than during independent learning^3–6. This phenomenon, known as interbrain synchronization (IBS), has been shown to support the efficient exchange of information⁷. Specifically, IBS facilitates the transfer of information about events that are temporally and spatially distant⁸, with an optimal time by aligning signals from one individual with another’s most receptive period for effective communication and encoding⁷.

Recent advances in interpersonal neuroscience suggest that inter-brain activity during cooperative learning reflects dynamic alignment within a distributed fronto-temporo-parietal network supporting language, joint attention, and mentalizing⁹. Rather than arising from a unitary process, inter-brain dynamics emerge through the interaction of functional specialized systems. For example, the inferior frontal gyrus (IFG), a core language hub, enables syntactic processing and turn-taking during dialog¹⁰; the sensorimotor cortex (SMC) contributes to verbal coordination and action prediction, forming a scaffold for real-time coupling of speech and gesture¹¹; the middle temporal gyrus (MTG) supports semantic integration and narrative coherence, particularly in the context of abstract or complex content¹². These regions—predominantly left-lateralized in language-dominant populations—comprise the neural basis for the alignment of cognitive representations and the co-construction of meaning during cooperative learning^5,13. Investigating how inter-brain dynamics emerge across this network offers critical insight into the neurobiological mechanisms underlying social learning.

While IBS has been observed in learner-to-learner interactions, the question of who initiates this synchrony and the primary direction in which the information flows remains unclear. The direction of IBS, for instance, may be influenced by interaction roles, such as status (leader vs. follower)¹⁴ or gender¹⁵. Moreover, since interaction partners must continually update their understanding of each other’s inner states, intentions, motivations, and affect in order to anticipate behavior and adjust accordingly¹⁶, it is likely that the observed directional flow will evolve over time. Thus, understanding the direction of IBS and its dynamics can provide insights into the underlying neurocognitive mechanisms of effective cooperative learning^17–23.

In this study, we investigated the direction of IBS within a cooperative learning group with emergent leadership. Over 30 years ago, Lord, De Vader, and Alliger recognized emergent leadership as “a major component of the social fabric of many organizations”²⁴. Indeed, evidence shows that groups allowing leaders to naturally emerge and orchestrate team processes can surpass those that do not²⁵. Recently, an extensive examination of research on emergent leadership has been conducted, covering several decades of study²⁴. This review has established a clear and common definition of emergent leadership: the degree to which an individual, without formal status or authority, is perceived by others as exerting leaderlike influence. Although it is widely accepted that leaders typically guide the group and impact collective outcomes, the specifics of information flow within the group are more complex.

To effectively quantify IBS in these interactions, traditional algorithms, like wavelet transform coherence (WTC) and phase-locking value, have been widely used. Notwithstanding, they do not discern the specific direction of this activity; that is, they cannot directly distinguish whether the influence is from A to B or from B to A^7,26,27. Granger causality analysis (GCA) is particularly effective at revealing directional influences by estimating how well one individual’s neural activities can forward-predict another’s²⁸. Several previous studies have successfully used GCA to parse directional information flow between individuals^14,29,30. For instance, research using GCA has linked the direction of neural synchronization to leadership¹⁴. In the present study, we used GCA to investigate the direction of information flow during cooperative learning with emergent leadership.

To gain a fuller understanding of leader-follower dynamics, we complemented our neural analyses with behavioral measures of verbal communication. While GCA captures the directionality of neural coupling between leaders and followers, communication behavior reflects how learners contribute to the cooperative learning process in observable terms^31,32. In this study, we focused on two core behavioral indicators: utterance score, which quantifies the accuracy and task relevance of verbal contributions, and utterance duration, which captures the amount of time each learner spent speaking¹⁴.

Theories diverge on the nature of leadership. The social dominance hypothesis suggests that leadership arises to facilitate exchange or avoid the costs of repeated contest competition³³. According to this view, leaders guide group activities through top-down influence, using authority to ensure seamless communication and enhance productivity³⁴. A critical component of this leadership dynamic involves sensorimotor communication, whereby leaders proactively transmit structured verbal and non-verbal signals—such as timing cues, gestures, and vocal prosody—that scaffold joint action, interpersonal coordination, and expressive synchrony^35,36. Through these mechanisms, leaders help align individual contributions into a unified and coherent group output. Furthermore, the dominance exhibited by leaders is also reflected in their management of conversation. Drawing on psycholinguistics theories, leaders are likened to authors who craft messages with precision, while followers interpret these messages as readers would a well-defined text, successfully deducing the subject matter³⁷. This model of communication, often referred to as the “literary model”, reinforces the leader’s dominance by emphasizing their complete control over the dialog³⁷. In cooperative learning, this dominance is further evident as leaders direct the learning agenda, processes, and sequence¹. Moreover, they initiate more dialog, implying a largely one-way flow of information from leader to follower¹⁴. Empirical studies have provided support for the social dominance hypothesis by employing various methodologies to investigate the directionality of information flow within social interactions. For instance, in verbal discussions, leaders demonstrate higher leader-to-follower neural synchronization relative to the reverse direction¹⁴. This asymmetric pattern has also been observed in creative verbal interactions, where leaders play a central role in guiding idea generation and evaluation²⁰. Likewise, in nonverbal contexts such as guitar duets, directional connectivity analyses reveal that leaders predominantly drive the interaction, with stronger synchronization from leader to follower³⁸.

Alternatively, the social interdependence hypothesis emphasizes mutual dependency in maintaining leader-follower relations, leading to a more bidirectional exchange³⁹. From the perspective of social interdependence theory, leadership is not defined by hierarchical control but by the ability to support positive interdependence among group members^40,41. In bidirectional interactions, leaders are those who promote cooperative goal structures, facilitate reciprocal influence, and dynamically respond to others’ action to maintain mutual engagement and joint problem-solving^42,43. This hypothesis posits that for these relationships to be sustainable, reciprocity must exist; a leader offering resources without reciprocation is unsustainable⁴⁴. For instance, a leader could provide guidance and, in return, anticipate a sense of prestige and commitment from their followers (e.g., the service-for-prestige theory)⁴⁴. In terms of communication, this hypothesis moves beyond the “literary model” to a “cooperative model,” where both speaker and listener are actively engaged in an iterative process of message construction and interpretation⁴⁵. In this model, individuals collaborate, repair, expand on, or replace the linguistic elements until they reach a mutually accepted version, reflecting the mutual responsibility participants bear for understanding each utterance³⁷. In cooperative learning, high interdependence is created by positive goal interdependence, which exists when group members perceive that their goals are attainable only if all members are successful⁴⁶. This interdependence unites members into a dynamic whole, where changes in one member or subgroup affect the state of others, translating into structures that promote mutual goals, combined achievements, and rewards that foster shared commitment, such as coordinating efforts to avoid the loss of collective incentives⁴⁶. In support of this notion, previous research demonstrates that verbal communication relies on a bidirectional neural communication system, where both speaker and listener anticipate each other, leading to mutual modulation of neural synchronization⁴⁷. In cooperative learning scenarios, verbal exchanges are important, with students not only absorbing information but also serving as “active listeners” through their engagement and anticipation.

This study investigated whether and how GCA can be used to understand the directional information flow in cooperative learning with emergent leadership. During the experiment, dyads were recruited to collaboratively analyze Chinese ancient poems, with their brain activity recorded by functional near-infrared spectroscopy (fNIRS) simultaneously (i.e., fNIRS hyperscanning)⁴⁸. Leadership in each dyad was not pre-assigned. Rather, emergent leadership was identified through micro-coding by trained research assistants, who evaluated instances of leaderlike behavior, such as summarizing group progress or proposing next steps (e.g., “Now that we agree on this point, I think we can move on”). This leadership coding was conducted independently from the analysis of verbal communication. For the latter, we assessed two behavioral indicators: utterance score, which quantifies the accuracy and task relevance of verbal contributions, and utterance duration, which measures the total time each participant spent speaking. In parallel, we collected participants’ subjective ratings of their learning experience, including perceived engagement and interpersonal connection. Unlike utterance score and duration, which were coded by independent raters, these self-reported measures reflect each participant’s internal experience.

While previous fNIRS hyperscanning studies have applied GCA to examine directional inter-brain coupling, key questions remain regarding how such coupling dynamically evolves during real-time cooperative learning. In particular, little is known about how directional neural information flow unfolds over time, adapts to emergent social roles such as leader and follower, and relates to behavioral coordination and subjective experience. To address these gaps, the current study implemented GCA within a spatiotemporal framework that integrated both inter- and intra-brain measures, enabling us to map the fine-grained temporal dynamics of neural directionality during interaction. Rather than assuming static coupling, we aimed to characterize how interpersonal neural dynamics evolve over time and across functionally distinct cortical regions. By linking these dynamics to behavioral roles and self-reported engagement, we sought to provide a richer understanding of cooperative learning as a temporally structured and socially situated process.

Two potential hypotheses were considered regarding the direction of information flow. Based on the social dominance hypothesis, we anticipated that information flow during cooperative learning would be largely one-way, from leaders to followers. Alternatively, according to the social interdependence hypothesis, it is expected that bidirectional information flow would underline successful cooperative learning, likely involving different processes and engaging brain regions associated with social mentalizing and action planning^49,50. Furthermore, since cooperation unfolds over time, we explored the temporal evolution of GCA to test the development of cooperative learning. To complement inter-brain dynamics, we examined intra-brain activity, which captures individual self-regulation processes^51,52. Together, these measures provide a more comprehensive view of the neural mechanisms of cooperative learning. In addition to temporal dynamics, we considered whether directional information flow would vary across functionally distinct cortical regions. Drawing on prior research, we hypothesized that leader-to-follower influence would be more prominent in regions associated with semantic processing and discourse organization (such as MTG)¹⁸. In contrast, follower-to-leader flow was expected to emerge in areas involved in sensorimotor coordination and verbal anticipation (e.g., SMC), reflecting the follower’s role in monitoring and adapting to the leader’s communicative cues^50,53–55.

Results

To investigate the direction of information flow during cooperative learning, we recruited ninety college students who were randomly paired into 45 dyads. These participants were instructed to collaboratively analyze a Chinese poem. During the experiment, the two participants sat side-by-side while two portable LIGHTNIRS systems simultaneously recorded hemodynamic signals of each dyad, alongside a digital video camera recording their discussion. After the experiment, two trained research assistants independently viewed the video recordings to determine the emergent leaders (see Supplementary Table S1 for detailed coding schema). Additionally, two additional coders, who were blind to the leader selection, annotated the verbal communication in the recordings to assess the “utterance score” and “utterance duration” of each dyad member. To examine dyadic communication, we summed the utterance scores and utterance durations of both individuals within each dyad. This approach provides a measure of the total task-relevant verbal contributions (utterance score) and total speaking time (utterance duration) for the dyad as a whole. GCA was conducted to parse the direction of information flow and its temporal dynamics (for details see Methods, Fig. 1).

Fig. 1. A schematic representation for the study purpose and analysis plans.

Communication coding and Granger causality analysis were performed to yield overall and temporal dynamics of utterance duration/score and directionality of information flow, respectively. The associations among communicative behaviors, Granger causality, and subjective learning experience in leaders and followers were also parsed.

Communicative behaviors in cooperative learning

In all 45 dyads we analyzed, a leader naturally emerged during cooperative learning. Leaders consistently outperformed followers, with leaders showing higher utterance duration (p < 0.001, η² = 0.23; Fig. 2A) and utterance score (p < 0.001, η² = 0.53; Fig. 2B). Pre-test scores did not significantly differ between leaders and followers, suggesting that leadership emergence was not simply driven by prior topic knowledge (Supplementary analysis 1). To examine how communication evolved over time, we analyzed utterance duration and utterance score across six 1 min intervals, separately for leaders, followers, and dyads. Figure 2C displays utterance duration per-minute. Kruskal–Wallis tests revealed no significant main effect of interval on utterance duration for dyads (p = 0.07), leaders (p = 0.06), or followers (p = 0.61, Fig. 2C), indicating stable speaking patterns across the task. In contrast, Fig. 2D presents cumulative utterance scores, where each point reflects the total learning-relevant content contributed by the leader/follower/dyad up to that minute. The main effect of interval on cumulative utterance score was significant at the dyadic level (p < 0.001, η² = 0.50), with comparable effects observed for leaders (p < 0.001, η² = 0.44) and followers (p < 0.001, η² = 0.11), suggesting a progressive accumulation of interpretive contributions over the course of the discussion. At the dyadic level, post-hoc analyses showed significant differences between consecutive intervals up to the 4th interval (corrected ps = 0.003, 0.005, and 0.01 for the 1st–2nd, 2nd–3rd, and 3rd–4th intervals, respectively), while no significant differences were found between the 4th and 5th intervals (corrected p = 0.33), as well as between the 5 and 6th intervals (corrected p = 0.27). Dyadic utterance score increased during the first 4 min and then plateaued. The post-hoc analysis for leader utterance scores yielded the same pattern as observed at the dyadic level. For followers, post-hoc comparisons revealed no significant differences across intervals (corrected ps = 0.07, 0.40, 0.91, 0.99, and 0.95, respectively).

Fig. 2. Behavioral coding results.

Leaders (n = 45) showed greater utterance duration (A) and higher utterance scores (B) than followers (n = 45). C Utterance duration per 1 min interval, plotted separately for the dyad (gray), leader (red), and follower (blue); dyadic values reflect the sum of leader and follower speaking times (n = 45 dyads). D Cumulative utterance score across the task, showing total content-relevant contributions accumulated up to each minute made by the dyad (gray), leader (red), and follower (blue); dyadic scores reflect the sum of leader and follower scores (n = 45 dyads). Error bars denote standard errors of the mean, with individual data points provided. *p < 0.05, **p < 0.01, ***p < 0.001.

Bidirectional information flow was spatially dissociated

GCA was conducted on the time series of homogeneous channels to determine whether leaders, followers, or both dominated the information exchange. The Kruskal–Wallis (K-W) test on the pairwise-conditional Granger causalities (GC) showed that the mean GC at channel 6 (CH6) from leaders to followers (L2F) was significantly lower than from followers to leaders (F2L) (F2L, corrected p = 0.04, η² = 0.11; Fig. 3A). Conversely, at CH23, the mean GC for L2F was significantly higher than for F2L (corrected p = 0.04, η² = 0.09; Fig. 3B). Moreover, at both CH23 (L2F) and CH6 (F2L), GC values in real dyads significantly exceeded those observed in pseudo-dyads, based on permutation-derived null distributions (ps < 0.001). CH6 and CH23 were roughly located in the SMC and MTG cortex, respectively⁵⁶. These findings suggest that information flow during cooperative learning with emergent leadership is bidirectional and spatially distinct. To further test whether the observed GC patterns and their temporal dynamics were specific to CH23 and CH6, we compared these channels with other non-target channels within the same network. Similar rise-then-fall trends were observed across channels, but GC strength at CH23 (L2F) was significantly higher, whereas no such difference was found for CH6 (F2L). These results suggest that leader-to-follower influence may be more localized, while follower-to-leader influence may engage a broader network (see Supplementary analysis 2; Fig. S1).

Fig. 3. Spatial distribution of Granger causality (GC) in both directions.

A The mean GC at CH6 from leaders to followers (L2F) was significantly lower than from followers to leaders (F2L) (n = 45 dyads). B The mean GC at CH23 from L2F was significantly higher than from F2L (n = 45 dyads). CH6 and CH23 were roughly located in the sensorimotor cortex and middle temporal gyrus cortex, respectively⁵⁶. CH23 corresponds to the T3 location in the 10–20 international system in this study. Error bars denote standard errors of the mean. *p < 0.05.

As a robustness check, we analyzed the same dyads during a free conversation task. No directional GC asymmetry was observed in the task-relevant channels (CH6: corrected p = 0.62; CH23: corrected p = 0.80), suggesting that the bidirectional pattern identified during cooperative learning is not a general feature of conversation. To further examine whether such effects may occur elsewhere in the brain, we conducted exploratory comparisons across all 24 homogeneous channel pairs. No significant directional effects were found (all corrected ps ranged from 0.62 to 0.93). These results suggest that the observed GC asymmetries are more likely specific to the cooperative learning context than to social interaction broadly.

Bidirectional information flow was temporally alike

Building on our investigation into the directionality of information flow, we focus on the GC from CH6 in the F2L direction and from CH23 in the L2F direction. The mean GC at CH23 from L2F (Fig. 4A), and at CH6 from F2L (Fig. 4B), both peaked during the 4th minute interval. Leading up to this peak within the first 4 min, there was a rise in GC, followed by a general decrease after the peak. The temporal dynamics in both directions followed similar patterns.

Fig. 4. Temporal evolution of Granger causality (GC) in both directions.

A At CH23, the mean GC from leaders to followers (L2F) initially increased during the 1–4 min interval, reaching its peak at the 4 min mark, followed by a subsequent decline (n = 45 dyads). B Similarly, the mean GC at CH6, from followers to leaders (F2L), also showed an ascending trend from the beginning until the 4 min mark, where it reached its maximum before decreasing (n = 45 dyads). Solid and dash lines depict the comparison between real and pseudo dyads. Error bars denote standard errors of the mean. Individual data points are added for the real dyads. *p < 0.05, **p < 0.01, ***p < 0.001.

Our permutation tests showed that, for the direction L2F at CH23, mean GC in real (interacting) dyads was stronger compared to pseudo dyads for the 2nd, 3rd, 4th, and 6th intervals (corrected ps < 0.005). Likewise, in the direction from F2L at CH6, our results revealed a more pronounced GC in the real dyads than in the pseudo dyads during the 3rd and 4th intervals (corrected ps < 0.001), as well as during the 5th and 6th intervals (corrected ps < 0.05).

To complement the inter-brain analysis, we examined intra-brain activation trajectories over time at two task-relevant channels (CH23 and CH6). Leader activation at CH23 was higher in the early phase of the task, while follower activation at CH6 increased over time (Supplementary analysis 3; Fig. S2). These role- and region-specific dynamics suggest distinct patterns of neural engagement, with leaders playing a more proactive role early on and followers gradually increasing their involvement as the task progressed. Control analyses further confirmed the functional specificity of CH23, which showed significantly stronger directional influence compared to non-target regions (see Supplementary analysis 2; Fig. S1).

Inverse relationship between GC directions

Although both GC directions exhibited similar temporal trajectories across the task, we next examined whether their overall magnitudes were related. A Spearman correlation between GC values at CH23 (L2F) and CH6 (F2L), computed over the full 6 min session, revealed a significant negative association (r = –0.33, p = 0.03). This result remained robust after excluding one dyad identified as an outlier (r = –0.32, p = 0.03, Fig. 5). The same pattern held when using log-transformed GC values. These findings suggest that directional influence was not evenly distributed across dyads: stronger GC in one direction was generally associated with weaker GC in the opposite direction. This inverse relationship supports the interpretation that cooperative learning involves structured, asymmetrical coordination strategies shaped by task roles, rather than purely reciprocal information flow. We further explored whether the degree of leader–follower role separation influenced behavioral or neural measures, but found no significant results (Supplementary analysis 4).

Fig. 5. Inverse relationship between GC directions.

Negative correlation between GC at CH23 (leader-to-follower) and CH6 (follower-to-leader) (n = 45 dyads).

Linking to communicative behavior and learning experience

We conducted correlation analyses to examine associations between directional GC and both behavioral and subjective measures of interaction, aiming to explore how neural dynamics relate to communication during cooperative learning. A positive association was found between the GC from L2F at CH23 and the leaders’ utterance score (r = 0.35, uncorrected p = 0.02), which remained significant after excluding one dyad as an outlier whose GC exceeded three standard deviations from the mean (r = 0.31, uncorrected p = 0.04, Fig. 6A). Furthermore, we found that the GC from F2L at CH6 correlated with the follower’s perceived peer support (r = 0.38, uncorrected p = 0.01, Fig. 6B). However, neither effect survived correction for multiple comparisons. No other associations—whether with utterance duration or subjective ratings such as similarity, liking, closeness, or engagement—were significant before or after correction (see Supplementary Table S2 for full results). To further test the robustness of these findings, we repeated the correlation analyses using log-transformed GC values to account for potential skewness. The CH23 association with leader utterance score remained significant, while no additional associations emerged. Full results are provided in Supplementary analysis 5. We also examined the associations between GC and Subjective Interaction Quality (averaged across five self-reported dimensions: liking, closeness, engagement, similarity, and support) and turn-taking dynamics (including turn frequency, leader-initiated turns, and follower-initiated turns), but found no significant effects (see Supplementary analyses 6 and 7).

Fig. 6. Associations among Granger causality (GC), utterance score, and perceived peer support.

A The GC from leaders to followers (L2F) at channel 23 (CH23) was positively correlated with the leaders’utterance score (n = 45 dyads). B The GC from followers to leaders (F2L) at channel 6 (CH6) was positively associated with the followers’ perceived peer support (n = 45 dyads).

Discussion

This study investigated the direction of information flow in cooperative learning with emergent leadership. Leveraging fNIRS hyperscanning and GCA, we made three important empirical contributions. First, we identified a bidirectional exchange of information between two learners, with notable L2F causalities in the left MTG and prominent F2L causalities in the left SMC. This finding highlights a key human trait: the capacity to transfer knowledge reciprocally between individuals³⁰. It underscores that this transfer is reciprocal and that distinct social roles, such as leaders and followers, involve varying coordination demands. Echoing prior research findings, leaders typically contribute more interactions, as evidenced by their significantly greater verbal output and richer informational content compared to followers^14,18,22. This proactive role underscores the leader’s responsibility in initiating discussions and guiding group dynamics, demonstrated by the positive correlation between leaders’ utterance score and the GC from L2F. Conversely, followers also exert considerable influence. The stronger GC from F2L indicates their active engagement, which, associated with their perceptions of support, completes the feedback loop of influence. To examine whether the observed bidirectional GC pattern is specific to cooperative learning, we conducted a control analysis in which the same dyads engaged in unstructured free conversation (see Results; see also Supplementary analysis 8). While this condition does not replicate key features of cooperative learning—such as shared goals and clearly emergent leader–follower dynamics—it provides a meaningful point of comparison. In this context, no significant GC directionality was observed, despite natural variation in speaking time between participants. Moreover, utterance duration was not significantly associated with GC values (rs < 0.25, ps > 0.10; full statistics provided in Supplementary Table S2), and additional indicators of verbal coordination—such as turn frequency and leader-/follower-initiated turns—also showed no significant associations with GC (see Supplementary analysis 6). These findings suggest that the observed directional information flow cannot be attributed solely to differences in speech quantity or surface-level interactional dynamics. Instead, the emergence of bidirectional GC appears to depend on the structured, cognitively demanding nature of cooperative learning, in which reciprocal engagement is shaped by shared objectives and asymmetrical roles. These findings support the view that bidirectionality is a distinctive feature of cooperative learning, rather than a generic consequence of imbalanced conversational dynamics, and highlight the need for further investigation into information flow patterns in other interaction contexts characterized by speaker–listener asymmetry.

Second, GC in both directions engaged different brain regions, suggesting a spatial hierarchy within this bidirectional communication. Specifically, the left MTG was associated with L2F information flow, while the SMC was primarily associated with F2L information flow. These findings align with previous research indicating that neural mechanisms during social interaction involve not only shared sensorimotor input but also shared mental processes, such as the representation of semantic and social concepts⁵³. Importantly, the neural mechanisms associated with shared linguistic components can be dissociated from those with shared sensorimotor input in anatomical locations⁵⁷. However, questions remain about the relationship between sensorimotor processes, internal mental processes, and social interaction, including their functional significance⁵³. In our study, leaders consistently exhibited longer and higher-quality utterances, often initiating proposals or summarizing the group’s interpretations, patterns suggestive of a more proactive, content-organizing role. This aligns with previous findings implicating the MTG in semantic integration, narrative construction, and the organization of meaning during communication^12,58. Notably, stronger follower-to-leader influence in the SMC was associated with followers’ perceived peer support. This pattern may suggest that when followers feel more supported, they engage more spontaneously through subtle verbalizations, gestures, or facial cues, which in turn elicit sensitive or responsive neural activity in the leader. The SMC has been shown to support turn-taking, speech monitoring, and anticipatory processing during interactive tasks^50,53–55, suggesting that enhanced F2L GC may reflect leaders’ sensitivity to followers’ spontaneous (perhaps independent or sporadic) motor actions. Taken together, these observations suggest that inter-brain information flow is modulated by participants’ functional roles within the interaction. The spatial dissociation between MTG and SMC is therefore not merely anatomical, but reflects differentiated contributions to cooperative learning behavior—semantic structuring in leaders and response coordination in followers—realized through distinct cortical pathways. The observed directional coupling may reflect a distributed, role-sensitive system supporting adaptive alignment in goal-directed cooperative learning.

Third, our analysis not only clarified the spatial distribution but also revealed the temporal evolution of bidirectional information flow. GC, as a proxy of information flow, increased initially, followed by a subsequent decline over the course of the cooperative learning process. By continuously tracking subtle changes in activity throughout the entire neural streams of cooperative learning, our findings indicate that synchronization involves more than merely “going on and off together” at the beginning and end of the learning session³⁰. Instead, natural social interactions are dynamically coupled, comprising complex sequences of interactions, reactions, and adaptations that evolve over time¹⁷.

The main theoretical contribution of this paper is an enhancement to the social interdependence theory—a long-standing theory in the field of cooperative learning that explains how cooperative learning strategies work—by providing a more specific delineation of the neural mechanisms. To achieve this, we examined neural interdependence and observed a bidirectional information flow that aligns with the principles of the social interdependence hypothesis.

The concept of interdependence within groups has been recognized for decades. In the early 1900s, Kurt Koffka proposed that groups were dynamic wholes in which the interdependence among members could vary; later, in the 1930s, Kurt Lewin identified common goals as the core that unites members into a dynamic whole, where individual or subgroup changes modify the whole; David Johnson and Roger Johnson further extended Lewin’s reasoning about interdependence and formulated a theory of cooperation, competition, and individualistic efforts⁴⁶. They proposed that cooperation arises through positive goal interdependence, where members believe their success is contingent on the success of others.

These theories have provided clarity on how cooperation functions within groups. However, the challenge remains to characterize the underlying mechanisms behind the theory. Interpersonal neuroscience offers a framework to articulate the theory and its underlying rationale⁵⁹. When cooperation is the norm, high levels of synchronization are linked to the group’s learning outcomes (the shared goal)^5,48. Our findings support the notion that group members reciprocally influence one another, further highlighting the importance of neural interdependence in cooperative learning settings.

Importantly, our findings suggest the potential existence of temporally structured patterns within cooperative learning, echoing prior work indicating that interpersonal neural signatures can delineate specific learning stages^17,57. Specifically, mean GC in both directions peaked at the 4th minute interval—showing an increase before the 4th minute of the task, followed by a decline. Such “rise-then-fall” patterns may reflect the natural trajectory of engagement in short-term cooperation—participants may initially devote greater cognitive and communicative effort to establishing shared understanding, with a subsequent decline possibly indicating task familiarity, reduced novelty, or cognitive fatigue. This interpretation is supported by the behavioral results, where utterance duration remained stable but utterance score plateaued in the later intervals, suggesting a tapering of new, task-relevant content. Complementing this, intra-brain activation at task-relevant regions revealed distinct temporal profiles across roles and cortical areas. At CH23 (MTG), both leaders and followers showed elevated activation during the first interval, followed by a general decline. This pattern may reflect early semantic integration and content organization demands during the initial phase of meaning negotiation. In contrast, at CH6 (SMC), follower activation gradually increased over time, suggesting growing engagement in verbal coordination and real-time adaptation—a pattern that aligns with their functional role and the observed follower-to-leader GC. Leader activation at CH6 was more prominent during the early to middle phases, possibly reflecting the demands of initiating and sustaining verbal output. Importantly, we do not suggest that these temporal dynamics are fixed. In longer interactions, such as hour-long conversations or semester-long group collaborations, the “rise-then-fall” pattern may emerge later, repeat in cycles, or be punctuated by additional phases (e.g., role shifts or renegotiation) as task demands and group cohesion evolve^32,60,61. Future studies are needed to test whether the temporal hierarchy we observed reflects a general organizational principle of cooperative learning.

Beyond replicating prior findings of inter-brain synchrony during cooperation, our study contributes novel insights into the temporal organization and role-dependent dynamics of directional neural coupling. Although GCA has been previously used in hyperscanning, our integration of GCA within a spatiotemporally resolved framework, together with behavioral coding and subjective reports, allowed us to conduct a more fine-grained examination of how neural information flow evolves during interaction. Rather than reflecting a fixed or static feature of cooperation, directional coupling fluctuated over time, with varying strength and spatial distribution depending on communicative role and cortical region. Notably, this perspective was not imposed a priori, but rather emerged from the fine-grained temporal analysis, demonstrating how methodological innovation can yield theory-generating insights into the neural basis of cooperative learning.

Despite its strengths, this study has several limitations. First, this study examined spontaneously emerged leaders, where there might be minimal power differences between learners within dyads. Exploring pre-assigned leaders within hierarchical relationships, particularly within the cultural context of China, could potentially reveal unidirectional patterns of information flow. Second, our experiments were conducted in a laboratory environment rather than a real classroom, raising questions about the transferability of our results to real-world educational environments. To enhance the practical relevance of our research, future studies could replicate these experiments in real-world classrooms⁶². Third, in smaller groups, individual performance can be heavily influenced by dominant figures, such as group leaders, and group size may impact the quality of leader-follower interactions and overall task performance⁶³. Future research should investigate the dynamics between leaders and followers in larger group settings. Moreover, although our findings could inspire further research into related topics in diverse environments, this study only concentrated on short-term cooperative learning and stranger-dyad cooperative learning. The generalization of this study requires further validation. Moreover, our work merely established the statistical relationship concerning information flow without causal inference^64,65. Future investigations will be required to consolidate this finding and explore the causation using e.g., mathematical modeling. Lastly, it is undeniable that the exclusion of right-hemisphere data limits the interpretation of our findings. Regions such as the right TPJ or right IFG, which are implicated in affective processing and broader social cognition, may also contribute to cooperative learning but could not be examined in this study.

Conclusion

In summary, while cooperation holds a crucial position in human learning, its neural mechanisms remain inadequately understood. Using fNIRS hyperscanning and GCA in a naturalistic cooperative learning scenario, this study provides evidence of the bidirectional flow of neural information. This flow exhibits spatial dissociation but temporal similarity, showcasing a hierarchical structure of bidirectional communication within cooperative learning.

Methods

Participants

Ninety college students (46 females, age: 20.83 ± 2.60 years old) took part in this study. All participants were right-handed and had no history of neurologic or psychiatric disorders. They were randomly assigned to form dyads, with each dyad consisting of members of the same sex who were complete strangers to each other. We tested only same-gender participants in order to mitigate inter-individual and inter-dyad variability^66,67. Written informed consent was obtained from all participants before the experiment, and each participant received 40 yuan as compensation. The experimental procedures were approved by the ethics committee at Zhejiang University (No. 2022009). All ethical regulations relevant to human research participants were followed. This study involved secondary analyses of data originally collected for different research purposes^68,69. Whereas the previous work examined movement synchrony and undirected interbrain synchrony, the current study focuses on directional information flow using Granger causality and integrates it with coded verbal behaviors and subjective learning measures to address a distinct set of research questions.

Experimental tasks and procedures

We used an adapted poetry learning task to evaluate the mastery of two Chinese ancient poems, “Bu Suan Zi · Ode to the Plum Blossom (卜算子·咏梅)” and “Bu Suan Zi · Huangzhou Dinghui Courtyard Residence (卜算子·黄州定慧院寓居作)”, selected for their comparable difficulty levels. The material was selected because these ancient poems are commonly included in Chinese school curricula, making them particularly engaging for university students. Each poem was accompanied by a set of three short-answer questions designed to assess different aspects of comprehension and literary analysis: (1) Identify key images [or “Yi Xiang (意象)” in Chinese]; (2) Find the most important verse and interpret the subjective thoughts and feelings conveyed by the author; (3) Recognize and analyze the type of rhetoric used in specific verses. This problem-solving task and its associated materials were previously used and validated in our recent work^5,68. The three sets of questions exemplify assessments commonly seen in classic national standard textbook and are designed to provide a moderate yet meaningful cognitive challenge. By requiring participants to collectively analyze and evaluate both the meanings and significance of poems, the task fosters active engagement, perspective-sharing, and the refinement of understanding – hallmarks of the cooperative learning process.

The experiment was conducted in a simulated classroom environment, which entailed an independent learning session (baseline) and a cooperative learning session. In the baseline session, participants analyzed poem independently within 6 min and then answered the questions individually. In the cooperative learning session, dyads of participants jointly analyzed the poem for 6 min. By the end of cooperative learning, participants had to select a representative to provide their answers. The order of the learning conditions was counterbalanced across participants. In this study, data from the baseline were used to normalize the data from cooperative learning in both GCA and intra-brain activity. To exclude the possibility that the results observed in cooperative learning session simply arose as a result of oral communication, we additionally collected data from participants freely discussed on the topic of their college experiences (a free conversation). During the experiment, the two participants sat side-by-side, with a digital video camera (Canon, G7X, Canon Inc., Tokyo, Japan) positioned directly in front of each dyad to record their discussion.

Determination of the emergent leaders

In our study, we did not pre-assign leadership roles to participants. After the experiment, two trained research assistants independently viewed the video recordings of the participants’ discussions. They were asked to ascertain the naturally emerging leaders within each dyad, and determine the emergent leaders¹⁴. Specifically, in line with recent methodological guidelines on assessing emergent leadership²⁴, we developed a scale that combines items from various instruments (see Supplementary Table S1). The scale incorporated three approaches previously used to identify emergent leadership: first, an evaluation of each group member’s lateral influence within the learning group (e.g., “To what extent do you perceive this learner influencing another learner?”); second, an assessment of each member’s functional role within the group (e.g., coordinating tasks, motivating others, resolving conflicts); and third, a forced-choice method in which raters selected the emergent leader within each dyad (i.e., “Which learner do you think is the leader?”). The rating items were predetermined and served dual purposes: as evaluation criteria for leadership behaviors and as a structured framework to guide forced-choice decisions. To ensure consistency across coders, raters underwent training using a detailed manual that included both criteria and illustrative examples. Supplementary Table S1 presents the full set of items and example behaviors (e.g., the coordination statement “Now that we agree on this point, I think we can move on”), adapted from established coding frameworks^14,20,70. Following the development of the scale and completion of rater training, two assistants independently rated each participant and identified the emergent leader within each dyad. Interrater agreement was achieved in 41 of the 45 dyads; discrepancies in the remaining cases were resolved through joint review and consensus discussion. To assess the reliability of the scale, we calculated the intraclass correlation coefficient (ICC) for the full scale (ICC = 0.79, p < 0.001), as well as separate ICCs for each subscale (lateral influence: ICC = 0.70, p < 0.001; functional roles: ICC = 0.69, p < 0.001). The internal consistency of the full scale was also acceptable (Cronbach’s α = 0.90). In all dyads, the individual identified as the leader through consensus also received the highest mean ratings on both subscales, supporting the internal coherence and robustness of the leadership classification procedure.

Coding of communicative behaviors

Two additional coders, who were blind to the leader selection, annotated the verbal communication in the recordings. We recruited new coders to avoid potential bias from the leader selection influencing the coding of behavior. For each dyad member, the coders assessed the “utterance score” and “utterance duration” of communication. Specifically, “utterance score” refers to the scoring points that appear in discussions. Coders were provided with a reference answer and explicit scoring criteria prior to the evaluation (see Supplementary Table S3). Points were allocated for responses that were both accurate and undisputed by peers, with distinctions made between points attributed to leaders and followers (Fig. 1).

“Utterance duration” refers to the duration of communicative behavior. When viewing the videos, the coders first identified instances of speech, whether it was continuous speech or speech interrupted by pauses shorter than 300 ms (i.e., pauses shorter than this threshold were treated as part of the same utterance)⁷¹. Non-speech sounds were not included. Then, the coders annotated the duration of each identified speech segment. Finally, they calculated the total duration of speech for each individual.

The intercoder reliability (based on ICC) was 0.82 for “utterance score” and 0.87 for “utterance duration.” Communicative behaviors were coded separately for leaders and followers in terms of utterance score (content relevance and expressiveness) and utterance duration (speaking time). To construct dyad-level metrics, we summed each member’s utterance scores and durations within each 1 min interval, treating the dyad as the unit of analysis. This summation approach captures the full communicative output of both individuals and aligns with prior work on group-level cooperation and emergent leadership^14,18. The 6 min task was segmented into six 1 min intervals, balancing temporal resolution with interpretability^70,72,73. Utterance duration was analyzed minute-by-minute to capture temporal fluctuations in verbal activity, while utterance score was modeled cumulatively to reflect the ongoing accumulation of task-relevant contributions. This approach allowed us to distinguish between momentary shifts in speech and the buildup of interpretive substance over time. Our analysis covered the entire cooperative learning period as well as distinct 60-second intervals within that period⁷³.

fNIRS data acquisition

Hemodynamic signals of participants in each dyad were simultaneously recorded by two portable LIGHTNIRS systems (Shimadzu Co., Japan). Each participant wore a “4 × 4” optode probe set, configured to measure 24 channels (eight sources and eight detectors, inter-optode distance = 30 mm). The probe set covered the left fronto-temporo-parietal regions. This decision was guided by previous findings indicating that language-mediated cooperation and inter-brain synchrony are most reliably observed in left-lateralized areas, including the left IFG, SMC, and MTG. Given the limited number of our fNIRS optodes, we prioritized left-lateralized region to ensure optimal coverage of the neural circuits most relevant to verbal coordination and social-cognitive processing in cooperative learning. The T3 and Cz locations in the 10–20 international system were used as reference points to aid in positioning. These brain regions have been previously associated with social learning and interactions in educational settings using fNIRS hyperscanning^{57,67,74–77}. Each source emitted light at three wavelengths (780, 805, and 830 nm), and raw intensity signals were recorded at a sampling frequency of 7.407 Hz. The correspondence between the NIRS channels and measurement points on the cerebral cortex was determined using the virtual registration method implemented in the NFRI toolbox⁷⁸, which allows for probabilistic mapping of optode locations to standard brain coordinates based on anatomical atlases (e.g., MNI space). This method has been widely adopted in fNIRS research to approximate the cortical projection of measurement channels when individual structural MRI data are not available.

fNIRS data analysis

Preprocessing

FNIRS data preprocessing for each dyad was conducted using custom MATLAB codes and involved two main steps. First, we applied correlation-based signal improvement (CBSI) to eliminate head motion artifacts⁷⁹. The CBSI approach reduced noise by maximizing the negative correlation between the HbO and HbR signals, assuming it is close to −1. Since the HbR signal mirrors the HbO signal following CBSI correction, only the HbO data were included in the further analysis. This decision was further bolstered by many prior fNIRS studies, which have demonstrated that HbO concentration possesses a higher signal-to-noise ratio compared to HbR^18,22,48 and provides greater sensitivity as an indicator of regional cerebral blood flow in fNIRS measurements⁸⁰. Second, we used a wavelet-based denoising approach to remove global physiological noise, such as blood pressure and respiration artifacts⁸¹. The denoising process involved the automatic detection of time-frequency points contaminated by the global physiological noise using WTC. Subsequently, the fNIRS signal was decomposed using the wavelet transform, and the wavelet energy associated with the identified contaminated time-frequency points was suppressed. Approximately 2% of channels were removed due to excessive noise or flat signals identified through visual inspection. If at least 12 channels (≥50%) were classified as ‘‘noisy,’’ the fNIRS measurements for that dyad were excluded. Based on these criteria, no dyads were excluded.

Granger causality analysis

GCA was conducted to investigate the direction of neural information flow during cooperative learning. Multivariate Granger Causality Toolbox in MATLAB was used to estimate the magnitude of Granger causality (i.e., GC) between two time series. GC is a statistical estimation of how much one time series can be predicted by the history of another time series, relative to how much it can be predicted by its own previous history, presented as a log-likelihood ratio⁸². Our GCA was based on the preprocessed HbO times series, generated by each participant dyad over homogeneous channels during the cooperative learning periods. We converted the preprocessed signals into z-scores using the mean and the standard deviation of the signals recorded during the independent learning (baseline) session⁵. In our study, we focused on the information flow between the leaders and the followers (i.e., leader → follower, and follower → leader). Specifically, we computed the pairwise-conditional causalities of both directions, with each channel pair analyzed while conditioning on all signals from the non-target channels¹⁴.

Linking to communicative behaviors and subjective learning experience

Apart from the communicative behaviors coded, participants’ rating of their learning experience was collected after the experiment. This assessment involved rating their perceived partner similarity, liking, closeness, support, and engagement in the learning process using a 9-point Likert scale (1 = not very much, 9 = very much, see Supplementary Table S4). We used Spearman correlation analyses to examine whether GC was associated with communicative behavior (“utterance duration” and “utterance score”) and subjective learning experience in cooperative learning.

Statistics and reproducibility

All analyses were conducted using MATLAB 2022b (MathWorks Inc., Natick, MA). Significance level was set at p < 0.05. False-discovery-rate (FDR) correction was applied to account for multiple comparisons.

Communicative behaviors

Kruskal–Wallis tests were used to examine differences in communicative behaviors between leaders and followers. These tests were performed separately for utterance duration and score. Subsequently, Kruskal–Wallis tests with post-hoc analysis were used to examine dyadic communicative behaviors over time. These tests were also performed separately for utterance duration and score.

Granger causality

Kruskal–Wallis tests were conducted for each homogeneous channel (24 channels in total) to compare the magnitude of directional information flow between the two directions (L2F vs. F2L). To evaluate the presence of meaningful information flow, we implemented a permutation-based testing procedure. Specifically, we generated 1000 pseudo dyads by randomly re-pairing individuals from different original dyads (e.g., Learner A from dyad 1 paired with Learner B from dyad 2), ensuring that no pair had interacted. For each pseudo-dyad, GC values were computed separately for each channel-direction pair (e.g., CH23 for L2F, CH6 for F2L), yielding empirical null distributions for each pair. GC values from real dyads were then compared against these channel-specific null distributions using z-tests. To control for multiple comparisons, FDR correction was applied across relevant comparisons, with a significance threshold of p < 0.05.

To examine the temporal dynamics of directional influence, we focused on CH23 (L2F) and CH6 (F2L), which had been identified as task-relevant. For each 1 min segment across the session, we constructed segment-specific null distributions based on pseudo-dyad GC values at the corresponding time interval. Real dyad GC values were then tested against these segment-level null distributions using the same z-test and FDR correction procedure.

Associations among communicative behavior, learning experience, and GC

Correlation analyses were conducted to explore potential relationships between distinct dependent variables, including communicative behavior (“utterance duration” and “utterance score”), learning experience (perceived partner similarity, liking, closeness, support, and engagement in the learning process), and GC (L2F and F2L).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Y.L.: conceptualization, data curation, formal analysis, investigation, validation, visualization, writing—original draft, writing—review and editing; Y.W.: formal analysis; C.S.: investigation; F.D.: investigation; Y.P.: conceptualization, funding acquisition, investigation, methodology, project administration, resources, software, supervision, validation, visualization, writing—review and editing.

Peer review

Peer review information

Communications Biology thanks Takayuki Nozawa, Simone Shamay-Tsoory and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: Jasmine Pan and Aylin Bircan. A peer review file is available.

Data availability

The data are accessible at https://osf.io/t4a3h/⁸³. Statistical software used for data analysis is explicitly mentioned in the corresponding sections. This study was not preregistered.

Code availability

Custom code used for data preprocessing and analysis is available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interest.

Supplementary information

The online version contains supplementary material available at 10.1038/s42003-025-08445-6.