Multi-brain neurofeedback: what are we training for?

Abstract

Multi-brain neurofeedback offers new possibilities for guiding social interaction by capturing and modulating interpersonal neural dynamics in real time. We propose a hierarchical framework where neurofeedback targets shared sensory dynamics (signal), socio-cognitive processes (functional), or social outcomes (system). We highlight key methodological challenges and potential real-world therapeutic and pedagogical applications.

Multi-brain neurofeedback offers new possibilities for guiding social interaction by capturing and modulating interpersonal neural dynamics in real time. We propose a hierarchical framework where neurofeedback targets shared sensory dynamics (signal), socio-cognitive processes (functional), or social outcomes (system). We highlight key methodological challenges and potential real-world therapeutic and pedagogical applications.

From hyperscanning to hyperfeedback

A growing body of hyperscanning (see Glossary) research links key features of social communication to the extent to which neurophysiological activity becomes coupled across individuals: fluctuations in inter-brain synchrony are impacted by shared engagement and social behaviors and, critically, contribute to social learning and bonding ^{1, 2., 3., 4.}. Recent findings from our group and others suggest that providing multi-brain neurofeedback through visual, auditory, or haptic information about ongoing synchrony fluctuations may boost social outcomes (Figure 1A,B ^{3., 4., 5., 6.}).

Figure 1.

Multi-brain neurofeedback: implementation and hierarchical theoretical framework.

(A) Representative multi-brain/body feedback interfaces developed by our team over the past decade, engaging thousands of dyads and groups. Examples include: (i) Mutual Wave Machine [3], (ii) Mutual Brainwaves Lab (incl. collaboration with Bad Bunny and Residente, not pictured) ^{5., 6.}, (iii) Compatibility Racer, (iv) Harmonic Dissonance, (v) Measuring the Magic of Mutual Gaze (ft. Marina Abramović), (vi) XenBox (ft. Mike Gordon and Bob Weir, Grateful Dead). (B) A typical multi-brain neurofeedback implementation. During naturalistic social interaction, neural signals are simultaneously acquired from multiple individuals and transformed into indices of interbrain synchrony, which are fed back in real time to support regulation of coupled neural activity. For illustrative purposes, the figure depicts a visual interface; however, feedback interfaces can be auditory, tactile, robotic, etc. (A). (C) Hierarchical levels of multi-brain neurofeedback training. Training targets can be defined at multiple levels of organization.

The possibility of modulating interpersonal neural dynamics and making them tangible in real time opens several avenues for real-world applications. For example, it may be used as an aid in psychotherapy to identify and target often subtle interpersonal markers that may disrupt relationships; or in team training and collaborative contexts where shared attention and reciprocal engagement shape learning and decision-making. However, much remains unanswered (see Outstanding questions). In this forum, we build on decades of research in (social) perception and action to lay out a framework intended to serve as a guide toward engineering neurophysiological intervention pathways and user interaction parameters tailored to specific social training targets.

Outstanding questions.

What is the appropriate training target for a given social context—signal-level synchrony, functional-level social coordination, or system-level social processes—and how do these levels relate causally?

Which synchrony metrics or models (static vs. dynamic) are valid as real-time training signals across hierarchical levels, and what evidence supports their online use?

What timing and latency constraints limit effective feedback, and how do these constraints shape biologically meaningful temporal contingencies?

How can AI and computational models help transform interbrain signals into behaviorally meaningful, temporally precise feedback for training?

What ethical, normative, and agency-related issues arise when attempting to modulate social coordination at the neural level?

Alt-text: Outstanding questions

Hierarchical levels of multi-brain neurofeedback training

Socially relevant interbrain synchrony fluctuations arise across multiple interdependent timescales (Table 1 and see Outstanding questions). For example, oscillatory phase-resetting at the signal level may impact non-signal-level outcomes, as demonstrated by evidence from our group and others showing that learners exhibited better learning after undergoing interbrain modulation at 6 Hz (a rhythm related to pitch processing) with their instructors while learning songs [2]. Multi-brain neurofeedback research should accordingly distinguish hierarchically organized training targets at signal, functional, and system levels (Figure 1C) and account for outcome specificity by leveraging context-appropriate metrics (Table 1 and see Outstanding questions; [3]).

Table 1.

Diversity of interpersonal metrics at hierarchical levels

Interpersonal metric	Measurement	Timescale	Hierarchical level
Pearson correlation	EEG/fMRI/fNIRS	milliseconds→seconds	Signal
Coherence	EEG/fMRI/fNIRS	milliseconds→seconds	Signal
Wavelet Transform Coherence	fNIRS/fMRI	minutes	Functional
Weighted Phase Lag Index	EEG	seconds→minutes	Functional
Circular correlation coefficient	EEG/fNIRS	seconds→minutes	Functional
Interbrain representational similarity	EEG/fNIRS	seconds→minutes	Functional
Information-theoretic measures	fMRI/fNIRS/EEG	seconds→hours	Functional/system
Cross-correlation	Movements	seconds→minutes	Signal/functional
Gaussian Markov modeling	Linguistic dynamics	seconds→minutes	Signal/functional
Cross-recurrence quantification analysis	Physiological signals	seconds→hours	Signal/functional/system

EEG, electroencephalography; fMRI, functional magnetic resonance imaging; fNIRS, functional near-infrared spectroscopy.

Signal-level synchrony metrics primarily capture low-level neural alignment (e.g., interpersonal sensory–motor coupling), such as instantaneous phase coherence computed continuously using short overlapping windows (e.g., ~250 ms). The logic of this approach is straightforward: if neural synchrony is plastic, reinforcing moments of synchrony may shape interpersonal neural dynamics through implicit conditioning. Because oscillatory coupling unfolds rapidly, effective modulation requires fast and reactive feedback matching the temporal precision of the underlying neural rhythms. Signal-level training primarily concerns implicit, reinforcement-based sensorimotor learning.

Functional-level multi-brain neurofeedback, in turn, directly targets processes known to lead to positive social outcomes, such as joint attention or mutual prediction, shifting the focus from raw signal similarity to strengthening representational alignment. Examples include feedback rewarding synchrony in neural signatures that have been linked to optimal turn-taking dynamics, joint attention, or shared intentionality. Functional-level synchrony metrics often capture slower or more integrative dynamics. For example, when dyads follow similar strategies in a task that involves representational updating or predictive coding (e.g., inferring the partner’s attention or updating social predictions), synchrony in higher-order brain regions is observed [7]. Such interbrain representational similarity measures operate at longer integration windows (e.g., 1–10 s, with or without temporal lags).

At the system level, the target of multi-brain neurofeedback becomes the coordination and stability of the social system itself. The emphasis shifts to emergent interaction properties, such as problem-solving efficiency, the robustness of social coordination, or the maintenance of interpersonal relationships. Feedback may be based on aggregated indicators of system performance rather than instantaneous neural coupling. The central idea is that altering macroscopic interaction dynamics—such as network efficiency—can foster more stable and adaptive collective behavior and relationships. For example, interbrain global network efficiency has been shown to predict team coordination [8]. System-level training may promote dynamic group-level hyper-brain networks, whose training parameters can be inferred from multivariate or higher-order representations (>3 people per group), including network-based measures (e.g., hypergraph modularity) or information-theoretic measures (e.g., mutual information or directed information flow [9]), estimated over long temporal windows (>60 s). This level of training engages social strategic adjustment (e.g., adapting joint policies to optimize collective performance).

Feedback can be delivered through different interfaces (visual, auditory, haptic; see examples in Figure 1A) and either continuously or intermittently [10]. For example, our group has used continuous feedback rewarding moving-window synchrony with brighter visuals or faster robotic movements while participants engaged in freeform (non)verbal social interactions. In addition to positive social outcomes, dyads exhibited higher interbrain synchrony over time, suggesting that hyperfeedback can inform social learning, particularly for sensory-motor, signal-level dynamics ^{2., 3.}. In other work, participants experienced intermittent feedback summarizing nontimelocked neural similarities during their (preceding) interaction. Findings suggest a link between perceived group synchrony and trust in strangers (a system-level outcome). Critically, intermittent feedback helps circumvent a central interaction design challenge that is often glossed over: audiovisual neurofeedback interfaces may interfere with the very social dynamics they aim to support, as they tend to compete with cues from the partner. For example, a visual summary of synchrony dynamics post-session may support metacognitive reflection by participants or clinicians, particularly in applied settings such as psychotherapy or social skills training. Haptic, robotic, or immersive audiovisual interfaces are viable real-time solutions but tend to be more technically involved.

Computational and methodological challenges

Another interaction design challenge arises from the fact that real-time neurofeedback systems introduce unavoidable latencies due to signal acquisition and processing. Such delays can disrupt the tight temporal contingencies required for biological integration, exposing key differences between natural face-to-face interaction and digitally mediated social exchanges. Moreover, interbrain synchrony is not sustained over time but tends to organize into transient episodes of elevated coupling, often preceded by periods of slow, low-synchrony negotiation. These episodes may reflect gradual build-up processes that culminate in moments of mutual understanding, coordinated turn-taking, or shared flow. Accurate detection of such bursts requires adaptive time windows matched to the evolving temporal scale of interbrain synchrony [11].

Emerging predictive modeling approaches to interbrain dynamics represent promising but nascent research directions. For example, coupled nonlinear oscillator models (e.g., van der Pol systems) fitted to real-time neural data can forecast coupling strength evolution, while hybrid architectures combining recurrent neural networks with dynamical systems constraints may enable the prediction of coordination states [12].

Real-time data cleaning presents a further dilemma: while necessary offline, aggressive online artifact rejection may especially impact moments of physically active social engagement, effectively penalizing desired states and disrupting the feedback loop.

Signal-level synchrony is especially sensitive to such artifacts, as well as other epiphenomenal signals generated by shared inputs. To mitigate artifact-driven coupling, real-time pipelines may incorporate motion monitoring, control channels (e.g., peripheral physiology), and metrics less sensitive to zero-lag coupling, alongside task designs that dissociate shared movement from target neural processes. Functional-/system-level synchrony more directly indexes computational alignment, such as shared predictions or internal representations, making it less likely to be a byproduct of social interaction. Reinforcing synchrony without level-specific interpretation may amplify epiphenomenal signals.

Multi-brain neurofeedback is also susceptible to awareness- and expectation-driven influences. For example, the mere presence of a social biofeedback environment encourages participants to engage in synchrony-inducing behaviors [3]. Rigorous control conditions, such as sham feedback or yoked controls, are therefore essential for disentangling genuine neural effects from expectancy-driven behavioral changes.

Critically, nonneural physiological and behavioral measures may also serve as sources of social biofeedback. Examples include movement dynamics capturing interpersonal coordination (e.g., synchronized posture, gestures, or gaze); linguistic alignment (e.g., pitch, speech rate, lexico-semantic properties, and turn-taking dynamics); or physiological synchronization reflecting shared arousal and emotional states (e.g., coupling in heart rate variability, skin conductance, and respiration). Multi-brain neurofeedback is most informative when neural signals capture latent or anticipatory processes not fully observable in behavior, including covert attention, prediction error signals, and pre-motor or pre-verbal alignment. Behavioral and physiological measures remain complementary, but neurofeedback uniquely enables monitoring and modulation of internal states, particularly when behavioral outputs are delayed, ambiguous, or strategically controlled.

It is important to emphasize that synchrony may be maladaptive, including conformity, dominance, or reduced exploration. As illustrated by, for example, psychotherapeutic phenomena such as ‘rupture and repair’ and ‘aha moments’ in teaching or teamwork, multi-brain neurofeedback may be most effective when targeting dynamic transitions between coupling and decoupling [13], rather than emphasizing ‘high synchrony’ as an outcome.

Concluding remarks

Multi-brain neurofeedback extends hyperscanning from observation to intervention. Its theoretical and practical impact, however, depends on clearly specifying what is being trained. We argue that target synchrony metrics must be aligned with underlying neural computations and intended behavioral outcomes. Integrating real-time modeling, ecological task design, and rigorous experimental controls will be essential for developing a mechanistic science of social coordination.

Glossary

Continuous feedback

feedback delivered in real-time with minimal delay, typically required for shaping fast neural dynamics at the signal level.

Dynamical systems

a mathematical framework for describing how system states evolve over time, often used to model coordination dynamics and transitions between interaction states.

Hypergraph

a generalized network representation in which edges can connect more than two nodes, enabling the modeling of higher-order interactions within multi-brain systems.

Hyperscanning

the measurement of two or more brains simultaneously.

Information-theoretic measures

metrics such as information decomposition and information flow approaches that capture how information is jointly represented across multiple variables.

Interbrain representational similarity

a metric quantifying the similarity in neural response patterns across individuals, often used to index shared cognitive representations.

Interbrain synchrony

temporal dependency between neurophysiological signals from two or more individuals.

Intermittent feedback

feedback delivered at discrete intervals (e.g., every 30 s) or at the end of trials, typically suited for higher-level cognitive (functional-level) or system-level training.

Multi-brain neurofeedback

a closed-loop paradigm in which interbrain signals are computed in real time and fed back to interacting individuals to modulate their neural and behavioral coordination.

Oscillatory phase-resetting

a transient alignment of neural oscillation phases in response to salient social signals (e.g., gaze, gestures, or vocalizations), often considered a mechanism underlying rapid synchronization across brains.

Recurrent neural networks

a class of neural network models designed to process sequential and time-dependent data by maintaining internal states that evolve over time.

Social biofeedback

a framework in which neural, physiological, or behavioral signals are fed back within a social interaction, enabling the regulation of both individual states and shared interpersonal dynamics.

van der Pol systems

nonlinear dynamical systems characterized by self-sustained oscillations and state-dependent damping, often used to model coupled oscillatory processes.