Computational theories of n-person conflict aim to explain triadic interactions between agents A, B, and C, in which C is drawn into dyadic conflicts between A and B in one of four scenarios: generalization, alliance, defense, and displacement (Pietraszewski, target article). Although Pietraszewski considered A and C forming a group in the displacement scenario, in which A attacks B, and B attacks C in a chain of attacks, we would argue that there is no group in this type, as the relationship between A and C is undetermined, which can be either (1) unrelated, (2) allied (so B attacks C to retaliate against A), or (3) indirectly antagonistic (A attacks B to cause B to attack C). Furthermore, we propose an additional scenario, namely compassionate mediation, in which C prevents A from attacking B in any of multiple ways, for example, C can deter A from attacking B by fortifying B's defense, or C can eliminate A's aggression by providing an alternative solution to A's problem that causes A to attack B. As the compassion here requires equanimity with friends and foes, with a genuine intention to relieve their suffering equally (Ho, Nakamura, & Swain, Reference Ho, Nakamura and Swain2020b), such compassion mediation is fostered by benevolent intersubjectivity, that is, attuning to self and other's needs with benevolence, which can be modeled in a dyadic active inference framework (Ho et al., Reference Ho, Muzik, Rosenblum, Morelen, Nakamura and Swain2020a), described below. We postulate that if C can first establish dyadic intersubjectivity with A and B separately, that is, C–A and C–B, and if C intends to use dyadic intersubjectivity to address A and B's intentions and/or needs equally, then C can accomplish the compassionate mediation between A and B (A–C–B). Although A and B would never be in the same group in Pietraszewski's conflict model, A, B, and C would form a group altogether as a result of compassionate mediation.

As the building blocks of compassion mediation, we summarize the dyadic active inference framework (Ho et al., Reference Ho, Muzik, Rosenblum, Morelen, Nakamura and Swain2020a) as follows: (1) Each agent has an active inference engine, in which there is an two-node interface of afferent feelings and efferent actions, along with an internal model node to infer the other agent's internal models; (2) if the internal model fails to reduce its prediction errors the agent will perceive stress; (3) when one agent's feelings are caused by the other's actions and vice versa, they are strongly coupled; (4) normally, such strong coupling enables the agents to reduce prediction errors of each other's internal models and thus reduce stress; (5) undercoupling will cause any agent to ignore the impact of their actions on another agent and thus results in failures to understand the other agent's internal model; (6) stress can cause over-mentalizing of others by preventing an agent from updating their internal model because of unresolvable conflicts; and (7) holding on to ineffective internal models despite failures to minimize prediction errors can result in a vicious cycle of over-mentalizing and undercoupling and exacerbate conflict.

Thus, modeling group conflict may include the active inference engine as cognitive primitives underlying the under-coupling and over-mentalizing problems described above. Here, we discuss a brain model of an active inference engine, based on the neural responses during maternal “mirroring” of the child's feelings and actions, which covaried with parenting stress reduction after an intersubjectivity-promoting parenting intervention (Ho et al., Reference Ho, Muzik, Rosenblum, Morelen, Nakamura and Swain2020a): When the mother mirrored the child's facial expressions, such valid and genuine mirroring deactivated the default-mode network (which would mediate the internal models of the mother's engine) and, conversely, activated the mirror neuron system (which would mediate the feeling/action interface of the mother's engine) and the salience network (which would be triggered by prediction errors).

The salience network plays a pivotal role in the postconflict adaptation. After a conflict triggers an agent's salience network, his or her compassionate mediation may be enabled if the agent can suspend the previous internal model in the default-mode network and engage in the mirror-neuron system to understand other agents more genuinely; otherwise, defensive reactions, rather than compassionate mediation, when the postconflict salience network activates the default-mode network, resulting in excessive perseverance of the failing internal model and thus over-mentalizing of others (Ho et al., Reference Ho, Nakamura and Swain2020b). The importance of the salience network after conflicts is corroborated as the impairment of the salience network is common to psychopathology, including substance use disorders (Goodkind et al., Reference Goodkind, Eickhoff, Oathes, Jiang, Chang, Jones-Hagata and Etkin2015).

The salience network overlaps with maternal behavior neurocircuit (MBN) that regulates the balance between aggression and care in the maternal brain (Swain & Ho, Reference Swain and Ho2017; Swain, Ho, Fox, Garry, & Brummelte, Reference Swain, Ho, Fox, Garry and Brummelte2019). Indeed, the MBN mediates sensitive parenting in infant development to becoming compassionate agents themselves (Ainsworth, Blehar, Waters, & Wall, Reference Ainsworth, Blehar, Waters and Wall1978; Elmadih et al., Reference Elmadih, Wan, Downey, Elliott, Swain and Abel2016; Kim, Strathearn, & Swain, Reference Kim, Strathearn and Swain2016; Kim et al., Reference Kim, Rigo, Leckman, Mayes, Cole, Feldman and Swain2015b; Mayes, Swain, & Leckman, Reference Mayes, Swain and Leckman2005). The MBN, thus, contains brain systems critical to conflict resolution (Eslinger et al., Reference Eslinger, Anders, Ballarini, Boutros, Krach, Mayer and Zahn2021; Guo, Moses-Kolko, Phillips, Swain, & Hipwell, Reference Guo, Moses-Kolko, Phillips, Swain and Hipwell2018; Hipwell, Guo, Phillips, Swain, & Moses-Kolko, Reference Hipwell, Guo, Phillips, Swain and Moses-Kolko2015; Swain, Reference Swain2011; Swain, Kim, & Ho, Reference Swain, Kim and Ho2011; Swain & Lorberbaum, Reference Swain and Lorberbaum2008; Swain, Lorberbaum, Kose, & Strathearn, Reference Swain, Lorberbaum, Kose and Strathearn2007), which can be modeled as adversely affected by psychosocial stressors and psychopathology (Ho & Swain, Reference Ho and Swain2017; Kim, Ho, Evans, Liberzon, & Swain, Reference Kim, Ho, Evans, Liberzon and Swain2015a; Moses-Kolko, Horner, Phillips, Hipwell, & Swain, Reference Moses-Kolko, Horner, Phillips, Hipwell and Swain2014; Pawluski, Swain, & Lonstein, Reference Pawluski, Swain and Lonstein2021; Swain et al., Reference Swain, Ho, Rosenblum, Morelen, Dayton and Muzik2017; Swain & Ho, Reference Swain and Ho2019, Reference Swain and Ho2021). Adaptive parent–child dyadic interactions and parent–parent–child or parent–child–child triadic interactions may shape the salience network (equivalent to MBN) in participating agents, such that they are more likely to employ compassionate mediation in the context of conflicts.

We hope to see a computational model that can explain all types of scenarios in which agent C may exert violent or non-violent interventions in the context of conflicts. Future computational models of conflict may consider a triadic active inference framework to explain agent C's participation in terms of how the agents' active inference engines are coupled with one another.