Aggression experience and observation drive shared behavioural andneural changes

A Nature study reports that mice trained through either direct aggression experience or observation of aggression develop a shared adaptive behavioral strategy absent in untrained animals, and that both routes produce shared neural changes in a core aggression network. Longitudinal cell-type-specific recordings across subcortical networks showed that observed aggression recruits widespread activation mimicking activity evoked by experienced behavior, and that these changes persist into a novel aggression context. Multiregion modeling identified a widespread reduction in inhibition across both experience and observation, which the authors found necessary for the adaptive behavioral change.

Abstract The ability to observe the social behaviour of others and use observed information to bias future action is a fundamental building block of social cognition<sup 1 ref-CR1 , 2 ref-CR2 , 3 https://www.nature.com/articles/s41586-026-11217-9 ref-CR3 </sup . Aggression observation activates hypothalamic neurons relevant for aggression control<sup 4 https://www.nature.com/articles/s41586-026-11217-9 ref-CR4 , 5 https://www.nature.com/articles/s41586-026-11217-9 ref-CR5 </sup , but whether observation and experience engage common circuit mechanisms that promote a new shared behavioural policy is unclear. Here we develop a strategy to perform large-scale cell-type-specific recordings across subcortical networks for social behaviour generation and learning. We record longitudinally while animals train through direct aggression experience or observation, then probe shared differences in behaviour and neural activity in a novel aggression context. After training, we detect unique signatures of a shared adaptive behavioural strategy that is absent in animals without training. During aggression observation, observed behaviour recruits widespread activation across subcortical networks for behavioural generation and mimics activity evoked by experienced behaviour, and shared neural changes persist into the novel aggression context. Multiregion modelling reveals shared changes to a core aggression network across both experience and observation, including a widespread reduction in inhibition, which we find is necessary for adaptive behavioural change. Together, this demonstrates that experience-like activity during aggression observation can recruit a shared inhibitory plasticity mechanism that biases behaviour toward adaptive proactive behavioural strategies. Similar content being viewed by others Main A major open question is whether experience and observation modify behaviour through a shared neural mechanism to generate common adaptive behavioural strategies. Models of social learning posit that social behaviours, like other behaviours, can be acquired or refined as a result of either direct experience or observation<sup 2 https://www.nature.com/articles/s41586-026-11217-9 ref-CR2 , 3 https://www.nature.com/articles/s41586-026-11217-9 ref-CR3 </sup . Social observation may be particularly important when the costs of learning through direct experience are high, as in aggression. Albert Bandura’s classic ‘Bobo doll’ experiments<sup 3 https://www.nature.com/articles/s41586-026-11217-9 ref-CR3 </sup , which demonstrated that children can learn aggressive behaviour through observation, highlighted that even behaviours that are considered part of an innate social repertoire can be highly influenced by social observation. In humans, exposure to violence is a risk factor that predicts an increased likelihood of future violence, suggesting that these effects can be long-lasting and generalize to new contexts<sup 6 https://www.nature.com/articles/s41586-026-11217-9 ref-CR6 , 7 https://www.nature.com/articles/s41586-026-11217-9 ref-CR7 </sup , but the underlying mechanism remains unclear. In animal models, chronic exposure to aggression has been shown to predict future aggression<sup 8 https://www.nature.com/articles/s41586-026-11217-9 ref-CR8 </sup . Moreover, in mice living in groups, where exposure to aggression is both frequent and behaviourally relevant, individuals attend to the fights of others and modulate their own aggression levels<sup 9 https://www.nature.com/articles/s41586-026-11217-9 ref-CR9 , 10 https://www.nature.com/articles/s41586-026-11217-9 ref-CR10 </sup . Previous work has identified a set of proactive behaviours, including self-defence and vigilance, that predict resilience and emerge through direct repeated aggressive social contact<sup 11 https://www.nature.com/articles/s41586-026-11217-9 ref-CR11 , 12 https://www.nature.com/articles/s41586-026-11217-9 ref-CR12 </sup . However, whether these adaptive behavioural strategies can also emerge after social observation and whether behavioural changes share a common neural mechanism is not known. Behaviour ‘mirror’ neurons have been described in cortical structures across a variety of species<sup 13 https://www.nature.com/articles/s41586-026-11217-9 ref-CR13 </sup , and recent evidence has shown that neurons in a key hypothalamic area for aggression<sup 14 https://www.nature.com/articles/s41586-026-11217-9 ref-CR14 </sup the ventromedial hypothalamus, ventrolateral area, VMHvl can also be recruited during vicarious observation of aggression<sup 4 https://www.nature.com/articles/s41586-026-11217-9 ref-CR4 , 5 https://www.nature.com/articles/s41586-026-11217-9 ref-CR5 </sup . However, whether this property is widespread within larger subcortical networks for social behaviour generation and whether this activity is critical for updating future behaviour in novel contexts is unclear. The social behaviour network SBN is a highly conserved network of largely subcortical brain regions hypothalamus, amygdala and midbrain that is conserved across vertebrates<sup 15 https://www.nature.com/articles/s41586-026-11217-9 ref-CR15 </sup . The SBN is highly interconnected, with many key regions in the core aggression network of the hypothalamus and amygdala sending and receiving both excitatory and inhibitory projection to and from other circuit loci<sup 16 https://www.nature.com/articles/s41586-026-11217-9 ref-CR16 , 17 https://www.nature.com/articles/s41586-026-11217-9 ref-CR17 </sup . Furthermore, this network is tightly connected with networks for appetitive and aversive learning that have been shown to promote and reinforce specific defensive strategies during naturalistic behaviour that lead to persistent affective states<sup 11 https://www.nature.com/articles/s41586-026-11217-9 ref-CR11 , 12 https://www.nature.com/articles/s41586-026-11217-9 ref-CR12 , 18 https://www.nature.com/articles/s41586-026-11217-9 ref-CR18 </sup . Although signals for appetitive learning, including dopamine release in the NAc, have been shown to respond vicariously to the actions of others in non-social contexts<sup 19 https://www.nature.com/articles/s41586-026-11217-9 ref-CR19 </sup , it is unclear whether these are activated by observing social actions during naturalistic social behaviours. While several previous studies have looked at the synaptic or hormonal basis of aggression experience within a subset of regions within the larger aggression network<sup 20 ref-CR20 , 21 ref-CR21 , 22 ref-CR22 , 23 https://www.nature.com/articles/s41586-026-11217-9 ref-CR23 </sup , how these changes influence large-scale neural coding across time or are replicated during aggression observation has not been explored. Previous work has implicated several inhibitory projections in the SBN as being critical for the gating and generation of aggressive motivational states and actions<sup 24 ref-CR24 , 25 ref-CR25 , 26 ref-CR26 , 27 ref-CR27 , 28 ref-CR28 , 29 https://www.nature.com/articles/s41586-026-11217-9 ref-CR29 </sup . One hypothesis is that inhibitory networks for aggression control exert less influence after aggression experiences, leading to new behavioural policies that favour aggression preparedness and other resilience-associated behaviours. Similarly, aggression observation could result in shared behaviour policies by co-opting changes to an inhibitory-control mechanism. Here we track the evolution of neural activity in a network for social behavioural generation and learning across experience. We find that both aggression experience and observation lead to a shared resilient behaviour strategy in a novel context that includes increased proactive aggressive and vigilance behaviours. Aggression observation evokes widespread activation across the SBN, with subsets of regions preferentially encoding different axes of observed behaviour. Moreover, we observe a widespread reduction in inhibition across the SBN, with these networks gradually becoming more biased towards excitation across either experience or observation. Importantly, these changes persist to novel contexts with observers exhibiting experience-like patterns of activation across the space of behaviour. Using a model-inspired perturbation strategy, we show that reactivation of key subcortical inhibitory regions after aggression observation reverts observers back to a naive, less-proactive behavioural state. Experience and observation shape behaviour To test whether aggression experience and observation similarly update aggression-control policies, we devised a behavioural train and test protocol Fig. 1a https://www.nature.com/articles/s41586-026-11217-9 Fig1 . Individual mice were first trained through direct experience or aggression observation or were provided with no previous exposure, then tested in a novel aggression context Supplementary Video 1 https://www.nature.com/articles/s41586-026-11217-9 MOESM4 . Aggression-experienced mice the EXP group were given opportunities to practice aggression in a safe and easy-to-win context 8 days, three 10 min sessions per day by giving them access to a novel, submissive mouse in their homecage. Mice that observed aggression the OBS group were placed onto a platform above similar aggressive interactions such that they could receive multimodal cues from exposure to this fight but could not interact with the behaving animals below. The non-social control group the NON group was given access to a mouse toy in their homecage for sessions of matched duration. After the 8 day training regime, all of the groups were then tested on an identical probe test: a novel aggression context with a more difficult opponent hereafter, hard fight . Hard fights comprised three brief 4 min interactions with a similarly sized mouse with previous aggression experience. To quantify behaviour across these train and test sessions, we used a combination of supervised classifiers, fine-grain aggression annotation and unsupervised clustering Fig. 1b https://www.nature.com/articles/s41586-026-11217-9 Fig1 . We tracked the pose of all individuals and converted pose to a series of individual and social behavioural features<sup 11 https://www.nature.com/articles/s41586-026-11217-9 ref-CR11 , 30 https://www.nature.com/articles/s41586-026-11217-9 ref-CR30 </sup . We trained a supervised behavioural classifier on tracked pose to detect attack frames across the training dataset Extended Data Fig. 1a https://www.nature.com/articles/s41586-026-11217-9 Fig6 and evaluated with a held-out dataset. We observed that, across the training epoch, the EXP and OBS groups experienced or observed many instances of aggression Fig. 1c https://www.nature.com/articles/s41586-026-11217-9 Fig1 at relatively stable rates across training in both groups Extended Data Fig. 1b–d https://www.nature.com/articles/s41586-026-11217-9 Fig6 . In the hard fight, while the total amount of attack did not significantly differ between the groups, the composition of that conflict was significantly different across groups Fig. 1d https://www.nature.com/articles/s41586-026-11217-9 Fig1 , with both the EXP and OBS groups showing increased initiated attacks with earlier attack latencies and fewer attacks received with lower latencies relative to the NON controls Fig. 1e–h https://www.nature.com/articles/s41586-026-11217-9 Fig1 . The NON controls, which neither observed nor experienced aggression, exhibited behaviours previously associated with stress susceptibility<sup 11 https://www.nature.com/articles/s41586-026-11217-9 ref-CR11 , 12 https://www.nature.com/articles/s41586-026-11217-9 ref-CR12 </sup , including more fleeing behaviour Extended Data Fig. 1e https://www.nature.com/articles/s41586-026-11217-9 Fig6 . These data show that both aggression experience and observation promote adaptive resilience-associated behaviours<sup 11 https://www.nature.com/articles/s41586-026-11217-9 ref-CR11 , 31 https://www.nature.com/articles/s41586-026-11217-9 ref-CR31 </sup , including more proactive aggression and resistance when attacked. To detect differences in behavioural strategy beyond attack, we performed unsupervised clustering of behaviour across training and the hard fight<sup 11 https://www.nature.com/articles/s41586-026-11217-9 ref-CR11 , 32 https://www.nature.com/articles/s41586-026-11217-9 ref-CR32 , 33 https://www.nature.com/articles/s41586-026-11217-9 ref-CR33 </sup Fig. 1i,j https://www.nature.com/articles/s41586-026-11217-9 Fig1 . To generate a common behavioural embedding, we performed nonlinear dimensionality reduction of the behavioural features, then performed density-based clustering of this low-dimensional behavioural embedding. Across all data, clusters could be mapped to interpretable behaviours Extended Data Fig. 2 https://www.nature.com/articles/s41586-026-11217-9 Fig7 , including a series of aggression preparatory postures, investigative behaviours, as well as distal vigilant and attentive states Fig. 1j https://www.nature.com/articles/s41586-026-11217-9 Fig1 . We found that a number of clusters defined by their low proximity between the animals contained attack behaviours and a number of aggression preparatory behaviours, while more distal clusters represented timepoints of social disengagement Supplementary Video 2 https://www.nature.com/articles/s41586-026-11217-9 MOESM5 and Extended Data Fig. 1f https://www.nature.com/articles/s41586-026-11217-9 Fig6 . We then projected subsets of data into this embedding by group and looked at differences in occupancy and transition structure Fig. 1k–o https://www.nature.com/articles/s41586-026-11217-9 Fig1 . Mapping the transition structure between clusters revealed a conflict supercluster, with an increased likelihood of transition between aggression and aggression preparatory behaviour Fig. 1m https://www.nature.com/articles/s41586-026-11217-9 Fig1 . To directly compare behavioural differences between the EXP, OBS and NON cohorts, we compared differences in both the cluster occupancy Fig. 1l https://www.nature.com/articles/s41586-026-11217-9 Fig1 and behaviour transition matrix Fig. 1m https://www.nature.com/articles/s41586-026-11217-9 Fig1 during the hard fight. We found significant group differences that are shared by the EXP and OBS groups relative to the NON control mice Fig. 1n,o https://www.nature.com/articles/s41586-026-11217-9 Fig1 and Extended Data Fig. 1g–j https://www.nature.com/articles/s41586-026-11217-9 Fig6 , indicating the emergence of shared changes to an aggression-control policy. Relative to the NON control mice, both the EXP and OBS mice spent significantly more time in several key aggression preparatory conflict clusters, which represent a gradient defined by velocity and body overlap Extended Data Fig. 2c https://www.nature.com/articles/s41586-026-11217-9 Fig7 , and significantly less time in inattentive asocial/avoidance clusters. We compared whether OBS behavioural strategies were more EXP-like or NON-like and found that the overall behavioural similarity between EXP and OBS was greater in cluster occupancy and behavioural transitions Fig. 1q–t https://www.nature.com/articles/s41586-026-11217-9 Fig1 and contained more shared axes of variance. Overall, these data suggest that both aggression experience and observation bias behaviour towards similar behavioural patterns in a novel aggression context. To test whether observation of aggressive behaviour was required for shared behavioural change, we exposed male mice to social but non-aggressive male mice the XPO group , then compared the behaviour of these socially exposed animals during the hard fight Fig. 1p https://www.nature.com/articles/s41586-026-11217-9 Fig1 . We found that the identity of the exposure group could be decoded from either the occupancy, persistence and transition of behaviour in the hard fight Extended Data Fig. 3a,b https://www.nature.com/articles/s41586-026-11217-9 Fig8 , and the behaviours performed by aggression observation and non-aggression exposure groups in the hard fight were distinct Extended Data Fig. 3c,d https://www.nature.com/articles/s41586-026-11217-9 Fig8 . Furthermore, hard fight behavioural patterns of XPO observation mice overlapped more with the behaviours of NON group mice Fig. 1p–t https://www.nature.com/articles/s41586-026-11217-9 Fig1 and Extended Data Fig. 3e,f https://www.nature.com/articles/s41586-026-11217-9 Fig8 than with the behaviours of OBS and EXP group mice. This indicates that observation of aggression is critical for experience-like behavioural change and exposure to non-aggressive behaviour in males is insufficient to produce adaptive behaviour change. Multiregion recordings during training We next explored whether changes in the activity of the SBN underlie shared behavioural changes across aggression experience and observation. We developed a two-colour multifibre recording strategy<sup 34 https://www.nature.com/articles/s41586-026-11217-9 ref-CR34 , 35 https://www.nature.com/articles/s41586-026-11217-9 ref-CR35 </sup that enabled us to record neural activity longitudinally at network-wide scale across training and the hard fight Fig. 2a https://www.nature.com/articles/s41586-026-11217-9 Fig2 . We targeted an extended SBN with custom thin-fibre arrays, including regions in the hypothalamus, amygdala, midbrain and forebrain Fig. 2b https://www.nature.com/articles/s41586-026-11217-9 Fig2 and Extended Data Fig. 4 https://www.nature.com/articles/s41586-026-11217-9 Fig9 . We also targeted two signals previously implicated in both social valence and appetitive and aversive learning during naturalistic aggression contexts: the lateral habenular LHb <sup 12 https://www.nature.com/articles/s41586-026-11217-9 ref-CR12 </sup and dopamine in the nucleus accumbens NAc-DA <sup 11 https://www.nature.com/articles/s41586-026-11217-9 ref-CR11 </sup . As previous work has implicated inhibitory projections in a core aggression circuit as a gating mechanism for aggression preparation and action<sup 17 https://www.nature.com/articles/s41586-026-11217-9 ref-CR17 , 25 https://www.nature.com/articles/s41586-026-11217-9 ref-CR25 , 26 https://www.nature.com/articles/s41586-026-11217-9 ref-CR26 , 36 https://www.nature.com/articles/s41586-026-11217-9 ref-CR36 </sup , we developed a strategy to map both putative excitatory and inhibitory signals simultaneously at each node in the network Fig. 2b https://www.nature.com/articles/s41586-026-11217-9 Fig2 and Supplementary Video 3 https://www.nature.com/articles/s41586-026-11217-9 MOESM6 , except for the NAc. We used a Vgat-ires-cre mouse to target putative inhibitory neurons Vgat<sup +</sup and non-inhibitory/putative excitatory populations Vgat<sup −</sup using a dual-colour viral strategy, and recorded dopamine release in the NAc using a genetically encoded dopamine sensor<sup 37 https://www.nature.com/articles/s41586-026-11217-9 ref-CR37 </sup . We used a custom multifibre isosbestic regression strategy for motion correction in the fighting animal Fig. 2c https://www.nature.com/articles/s41586-026-11217-9 Fig2 . We compared the signals across the EXP, OBS and NON groups during the training phase aligned to behaviours of interest Fig. 2d https://www.nature.com/articles/s41586-026-11217-9 Fig2 . NON control mice were aligned to yoked behavioural control timepoints from an EXP mouse. We observed robust activation across our neural populations in the EXP and OBS groups during training, and relatively little activity in the NON controls. Aggression observation activates the SBN To test whether observers were attentive to attacks, we trained a supervised classifier with a convolutional network on the spatiotemporal features of the raw video of the observer watching aggression during the training period Extended Data Fig. 5a,b https://www.nature.com/articles/s41586-026-11217-9 Fig10 . The classifier was trained to identify instances of attentive behaviours compared with other inattentive behaviours, including rearing, grooming and turning Fig. 2e https://www.nature.com/articles/s41586-026-11217-9 Fig2 and Extended Data Fig. 5c–e https://www.nature.com/articles/s41586-026-11217-9 Fig10 . We found that attention could be robustly predicted, and that observers increased the time spent in attentive behaviours across the training epoch Fig. 2f https://www.nature.com/articles/s41586-026-11217-9 Fig2 and Extended Data Fig. 5f https://www.nature.com/articles/s41586-026-11217-9 Fig10 and spent more time in attentive states relative to the XPO group Extended Data Fig. 5f,g https://www.nature.com/articles/s41586-026-11217-9 Fig10 . We observed that observers were more likely to attend during attack than during control timepoints and exhibited an increased probability that they were in an attention epoch after attack onset Fig. 2g https://www.nature.com/articles/s41586-026-11217-9 Fig2 and Extended Data Fig. 5i–k https://www.nature.com/articles/s41586-026-11217-9 Fig10 . Finally, we found that neural activity in observers was significantly more active during attended attack than during unattended attack Fig. 2h https://www.nature.com/articles/s41586-026-11217-9 Fig2 and Extended Data Fig. 5l,m https://www.nature.com/articles/s41586-026-11217-9 Fig10 . We next compared the response profiles during attack across the SBN between individuals experiencing and observing aggression. To compensate for trial-to-trial differences in shifting attention and for discrepancies in the onset of classifier aligned attack, we filtered inattentive trials and performed a network-wide time-shifting procedure<sup 38 https://www.nature.com/articles/s41586-026-11217-9 ref-CR38 </sup . This procedure fits a tensor-wide population-specific template and produces trial-to-trial shifts relative to classified attack using a narrow window Fig. 2i https://www.nature.com/articles/s41586-026-11217-9 Fig2 and Extended Data Fig. 6a https://www.nature.com/articles/s41586-026-11217-9 Fig11 . We identified a shared network-wide moment of activation followed by a sustained post-attack response in many regions across the SBN Fig. 2j https://www.nature.com/articles/s41586-026-11217-9 Fig2 and observed a significant sustained increase in activity aligned to attack in the vast majority of signals, even without time-shifting Extended Data Fig. 6b https://www.nature.com/articles/s41586-026-11217-9 Fig11 . We observed no significant attack-aligned activity in any signal in the NON mice, even after time-shifting Fig. 2j https://www.nature.com/articles/s41586-026-11217-9 Fig2 grey . Importantly, this procedure does not produce spurious response peaks aligned to synthetic attack timepoints, or matched behavioural timepoints of high-velocity movement without aggression Extended Data Fig. 6c,d https://www.nature.com/articles/s41586-026-11217-9 Fig11 . Overall, we observed an attack-aligned activity profile in the OBS mice that resembles the brainwide response in EXP mice, with sustained responses in a number of key inhibitory regions with known roles in aggression control, including the VMHvl, MeA, BNST and vLS<sup 24 ref-CR24 , 25 ref-CR25 , 26 ref-CR26 , 27 ref-CR27 , 28 ref-CR28 , 29 https://www.nature.com/articles/s41586-026-11217-9 ref-CR29 </sup . Decay constants fit to attack aligned responses in activated regions during observed aggression in the OBS group were significantly lower than the constants fit to aggression in the EXP group Extended Data Fig. 7a,b https://www.nature.com/articles/s41586-026-11217-9 Fig12 , indicating that the signal may not persist during observation to the same degree as during action. The decay time in inhibitory regions was slower than for corresponding excitatory signals Extended Data Fig. 7c,d https://www.nature.com/articles/s41586-026-11217-9 Fig12 , suggesting that persistence of inhibition is common to both experienced and observed aggression. Valence signals encode observed features Previous work has demonstrated that performing attack increases dopamine release in the NAc<sup 11 https://www.nature.com/articles/s41586-026-11217-9 ref-CR11 , 39 https://www.nature.com/articles/s41586-026-11217-9 ref-CR39 </sup , while being attacked increases the response of LHb excitatory neurons<sup 12 https://www.nature.com/articles/s41586-026-11217-9 ref-CR12 </sup . Consistent with this, we observed that the profiles of NAc-DA and LHb in EXP mice are highly biased towards NAc-DA Fig. 3a,b https://www.nature.com/articles/s41586-026-11217-9 Fig3 , both at the level of the individual and during simultaneously recorded signals at the single-trial level Fig. 3b https://www.nature.com/articles/s41586-026-11217-9 Fig3 . We observed significant activation of both NAc-DA and LHb during observed aggression Fig. 3a https://www.nature.com/articles/s41586-026-11217-9 Fig3 , although the response is not biased towards NAc-DA or LHb at the level of the individual or the single trial Fig. 3b https://www.nature.com/articles/s41586-026-11217-9 Fig3 . To test whether neural signals vicariously reflect unique features of observed attack that could account for trial-to-trial heterogeneity, we examined whether neural signals preferentially encoded actions and aggressor identity. We explored in an unbiased manner whether neural signals were differentially modulated by the actions performed by the aggressors Fig. 3c https://www.nature.com/articles/s41586-026-11217-9 Fig3 or the subordinate male Fig. 3d https://www.nature.com/articles/s41586-026-11217-9 Fig3 . We first compared neural activity defined by aggressor action the most vigorous high-velocity attacks compared with the least vigorous low-velocity attacks, Fig. 3c https://www.nature.com/articles/s41586-026-11217-9 Fig3 and tested whether individual signals would be sensitive to this difference in observed action. To quantify this, we trained a separate decoder for each neural population and rank-ordered the decoder accuracy performance. Second, we compared neural activity defined by subordinate action the fastest high-velocity escape trials compared with the slowest, low velocity trials; Fig. 3d https://www.nature.com/articles/s41586-026-11217-9 Fig3 . We found that NAc-DA was the only signal after multiple-comparison testing that reliably discriminated between both aggressor and subordinate action, suggesting a unique role in vicarious action representation. The LHb was not significantly modulated by observed action and performed no better than chance. These data demonstrate that appetitive but not aversive teaching signals encode information about observed actions of both individuals. To more directly test whether activity in observers vicariously reflected the behaviour of both individuals being observed, we recorded neural signals during attack where there was no subordinate individual, only an aggressor Fig. 3e https://www.nature.com/articles/s41586-026-11217-9 Fig3 . To perform this experiment, we used optogenetics to facilitate a simulated attack on an inanimate object, rather than a live subordinate conspecific. In the observed aggressor, we expressed ChR2 in the VMHvl of Esr1-cre male mice and implanted a stimulation fibre over the VMHvl. We recorded the neural signals of observers as they watched either conspecific aggression or stimulation-evoked attack on a toy mouse Fig. 3e https://www.nature.com/articles/s41586-026-11217-9 Fig3 , and trained a per-signal decoder to discriminate between real and simulated attacks. Similarly, we found that NAc-DA but not LHb was sensitive to the absence of the actions of subordinate individuals. Moreover, hypothalamic regions, including the VMHvl, PMv and AH, were significantly more robustly activated by the real attack than the simulated attack Fig. 3e https://www.nature.com/articles/s41586-026-11217-9 Fig3 . Finally, we tested whether neural signals across the SBN would be sensitive to other features of aggression beyond action. Recent work has demonstrated that neural activity in the MeA is modulated by familiarity in observed aggression<sup 40 https://www.nature.com/articles/s41586-026-11217-9 ref-CR40 </sup . To test whether other signals in the SBN would be sensitive to the identity of the observed individual, we generated familiarity by allowing the aggressor to attack the observer before observation Fig. 3f https://www.nature.com/articles/s41586-026-11217-9 Fig3 . We then recorded these observers while they watched either the familiar aggressor or a novel aggressor attack a novel subordinate mouse. We found that previous familiarity with the aggressor significantly amplified observed attack response and that familiarity is widely encoded across the SBN, including the LHb Fig. 3f https://www.nature.com/articles/s41586-026-11217-9 Fig3 . Notably, NAc-DA was not significantly modulated by familiarity, suggesting a unique role for this signal in the encoding of vicarious action, but not familiarity. Observation activity mimics experience To test whether neural activity during aggression observation mimics the response during experience across behaviours, we projected the activity of each signal from the training phase onto the behaviour map Fig. 4a https://www.nature.com/articles/s41586-026-11217-9 Fig4 and then compared the similarity across groups in this signal embedding Fig. 4b https://www.nature.com/articles/s41586-026-11217-9 Fig4 . We compared neural activity during the training phase Fig. 4c https://www.nature.com/articles/s41586-026-11217-9 Fig4 , across highly different behavioural conditions experiencing aggression, observing aggression and non-social control behaviour , and also in the hard fight Fig. 4d https://www.nature.com/articles/s41586-026-11217-9 Fig4 , in which mice performed similar types of behaviours. For each neural signal, we quantify the similarity between the EXP-to-OBS and OBS-to-NON groups using cosine distance to indicate the similarity of neural-behaviour maps across groups. We found that, across both excitatory and inhibitory signals in SBNs, the activity in OBS mice is significantly more similar to EXP mice than to NON mice, with hypothalamic inhibitory regions exhibiting the most similarity during both the training phase and during the hard fight Fig. 4e,f https://www.nature.com/articles/s41586-026-11217-9 Fig4 . We observed similar patterns by comparing neural responses tuned to individual behavioural features Extended Data Fig. 8 https://www.nature.com/articles/s41586-026-11217-9 Fig13 . These data indicate that, in addition to promoting experience-like behaviour, aggression observation can generate experience-like neural profiles at the brain-wide and behaviour-wide scale Fig. 4g https://www.nature.com/articles/s41586-026-11217-9 Fig4 and that changes persist into novel social contexts. Shared reduction in network inhibition We predicted that aggression experience and observation enabled similar adaptive behavioural strategies through a shared plasticity mechanism during training that is absent in the NON group. To test this, we compared neural activity during the hard fight and found both EXP and OBS individuals exhibited significantly reduced inhibition relative to the NON controls across a broad subset of recorded regions including the VMHvl, PMv, AH and vLS Fig. 5a https://www.nature.com/articles/s41586-026-11217-9 Fig5 . Notably, we did not observe significant differences across groups in any excitatory signals, suggesting that broad changes in network-wide inhibition, rather than excitation, underlie the capacity for differences in behavioural strategy. We observed a similar decrease in inhibition in neural signals aligned to conflict in the hard fight Extended Data Fig. 9a,b https://www.nature.com/articles/s41586-026-11217-9 Fig14 . To test whether these differences from NON mice reflect shared changes in inhibition, we trained a linear classifier to discriminate EXP from NON mice and projected OBS activity onto the learned axis Fig. 5b,c https://www.nature.com/articles/s41586-026-11217-9 Fig5 . Classifiers trained on inhibitory populations, but not excitatory populations, discriminated EXP from NON mice Fig. 5d https://www.nature.com/articles/s41586-026-11217-9 Fig5 . Projected onto the inhibitory axis, OBS mice differed from NON mice but not significantly from EXP mice Fig. 5e https://www.nature.com/articles/s41586-026-11217-9 Fig5 , with the largest contributions to OBS placement coming from a subset of inhibitory regions Fig. 5f https://www.nature.com/articles/s41586-026-11217-9 Fig5 . Overall, these results suggest that a shared reduction in inhibition may underlie the shared emergent experience-like behavioural strategy in individuals that observe aggression. We predicted that a shared reduction in inhibition in the SBN might result in a relaxation of inhibitory control to the VMHvl, a key hypothalamic site for aggression control<sup 14 https://www.nature.com/articles/s41586-026-11217-9 ref-CR14 , 24 https://www.nature.com/articles/s41586-026-11217-9 ref-CR24 , 25 https://www.nature.com/articles/s41586-026-11217-9 ref-CR25 </sup . To quantify this changing relationship, we modelled the VMHvl signal as a linear combination of ongoing activity from associated excitatory and inhibitory regions in the SBN Fig. 5g https://www.nature.com/articles/s41586-026-11217-9 Fig5 . We partitioned the contributions to the explained variance of the VMHvl into shared and unique variance using banded regularization<sup 41 https://www.nature.com/articles/s41586-026-11217-9 ref-CR41 </sup Fig. 5h https://www.nature.com/articles/s41586-026-11217-9 Fig5 . Using this approach, we trained a separate model for each day of training, including the hard fight. We observed a decrease in the unique variance explained by several inputs in the EXP models that is not present in the NON models Fig. 5i https://www.nature.com/articles/s41586-026-11217-9 Fig5 , indicating that, across time, the VMHvl activity is less influenced by specific inputs. Importantly, during the hard fight, we observed a similar significant drop in the unique variance of several inhibitory signals in the EXP group compared with in the NON group, including the inhibitory PMv, AH, MeA and vLS, which are shared by the OBS group Fig. 5j https://www.nature.com/articles/s41586-026-11217-9 Fig5 . Importantly, we did not observe a similar decrease in unique variance with excitatory signals during the hard fight. Reactivating inhibition reverses change Our modelling approach enabled us to identify a key set of inhibitory control regions that show changes in observers that mimic the changes in experienced individuals PMv, MeA, AH and vLS . We tested the causal contributions of this activity on changes in behavioural strategy after observation of aggression. We predicted that reactivation of these neural signals after observation would reverse the behavioural change induced by aggression observation. To perform the reactivation experiment, we expressed excitatory DREADD-G<sub q</sub <sup 42 https://www.nature.com/articles/s41586-026-11217-9 ref-CR42 </sup bilaterally in the suite of model-identified subcortical regions in Vgat<sup +</sup Cre mice and waited 3 weeks for viral expression Fig. 5k,l https://www.nature.com/articles/s41586-026-11217-9 Fig5 . We allowed these DREADD-G<sub q</sub OBS-G<sub q</sub mice and a matched control cohort OBS-control to observe aggression across the training phase, then tested behaviour in the hard fight. Both the OBS-G<sub q</sub and OBS-control cohorts received injections of synthetic ligand clozapine- N -oxide CNO 30 min before the hard fight, which reactivated suppressed inhibitory regions in the OBS-G<sub q</sub group but not in the control group. Using both supervised and unsupervised strategies for behavioural analysis in the hard fight, we found that the inhibition-reactivated OBS-G<sub q</sub cohort delivered significantly fewer attacks and received more attacks than the control cohort Fig. 5m https://www.nature.com/articles/s41586-026-11217-9 Fig5 , consistent with a reversion away from proactive social behaviours to more reactive aggressive strategies. Finally, both the distribution of occupancy across behavioural clusters and the structure of transitions between them was significantly more NON-like in OBS-G<sub q</sub mice than in OBS-control mice Fig. 5n,o https://www.nature.com/articles/s41586-026-11217-9 Fig5 , demonstrating a reversion back to a naive behavioural state. Overall, these data demonstrate the necessity of the inhibitory reduction in mediating adaptive behaviour changes after observation of aggression, and indicates a common mechanism by which observation can generate behavioural policy similar to aggression-experienced individuals Extended Data Fig. 10 https://www.nature.com/articles/s41586-026-11217-9 Fig15 . Discussion Here we show that both aggression experience and observation recruit a brain-wide network of excitatory, inhibitory and valence signals, and that changes to this activation pattern predict a shared adaptive behaviour-control policy in a new aggression context. First, both groups EXP and OBS show behavioural signatures of resilience, including increased initiated attacks and decreased latency to attack in the novel context—a classic signature of aggression instigation<sup 23 https://www.nature.com/articles/s41586-026-11217-9 ref-CR23 , 43 https://www.nature.com/articles/s41586-026-11217-9 ref-CR43 , 44 https://www.nature.com/articles/s41586-026-11217-9 ref-CR44 </sup . Both groups exhibited a new shared strategy consisting of two interpretable behavioural signatures: first, individuals in the EXP and OBS groups spend more time in aggression preparatory postures and similarly transition between aggression-preparatory clusters, indicating their readiness to fight. Second, both groups spend less time in distant inattentive states in which they might be vulnerable to attack. We interpret these behavioural changes as a common ‘be ready’ defensive strategy. Notably, behaviours that involve direct social contact show greater change after direct experience than after observation, suggesting that there may be some limits to what can be changed through observation alone. While aggression mirror neurons have been previously reported in a single hypothalamic population oestrogen- and progesterone-receptor-expressing VMHvl neurons, a subset of excitatory VMHvl neurons<sup 4 https://www.nature.com/articles/s41586-026-11217-9 ref-CR4 </sup , our data suggest that the representation of aggression during observation may be a network-wide phenomenon, activating multiple cell types across an extended aggression-control network. The current data are consistent with models of embodied social cognition, in which observation of aggression might activate neural populations associated with aggressive arousal, physiological signalling and reward monitoring, rather than merely encoding a reflection of an action or fear state. We found that NAc-DA, a signal known to be involved in encoding action value<sup 45 https://www.nature.com/articles/s41586-026-11217-9 ref-CR45 </sup and vicarious non-social reward<sup 19 https://www.nature.com/articles/s41586-026-11217-9 ref-CR19 </sup , vicariously responds to observed actions during naturalistic social behaviour and encodes differences in the behaviour of the aggressor and subordinate animal. Moreover, the response during a simulated attack is substantially reduced when the subordinate is absent. These data suggest that the observer may not automatically take the point of view of the aggressor and, instead, may learn vicariously from both the animal being attacked and also from the aggressor. This mix of perspectives is consistent with previous work using ‘witness defeat stress’ that shows that individuals exposed to aggression may develop long-term social policies and suggests that vicarious learning may emerge from attending to the actions of both the aggressor and the animal being attacked<sup 46 ref-CR46 , 47 ref-CR47 , 48 https://www.nature.com/articles/s41586-026-11217-9 ref-CR48 </sup . Rather than solely supporting a model of appetitive reinforcement during aggression, these data may suggest a broad form of goal emulation<sup 49 https://www.nature.com/articles/s41586-026-11217-9 ref-CR49 </sup , whereby the observer may learn a strategy avoid being attack-vulnerable that requires a vicarious reflection of outcomes from both participants. We find that a signal previously associated with aversive learning, the LHb, is not modulated by differences in observed actions and is instead amplified by familiarity with the observed individual. Changes in LHb signalling in bystanders observing a familiar conspecific in pain have also been observed<sup 50 https://www.nature.com/articles/s41586-026-11217-9 ref-CR50 </sup , suggesting that familiarity may modulate vicarious punishment signals. Across training, we reveal shared changes in network-level plasticity across the training phase that are common to both experienced and observed mice and are not present in the non-social group. These changes support a network-wide relaxation of inhibition during critical social interactions and demonstrate that these changes persist into a new aggression context. In particular, we observe a shared motif across the core hypothalamic aggression circuit, whereby inhibition becomes less predictive of neural signals in the VMHvl. Here we find direct circuit-level evidence for the decreased influence of inhibition in aggression-control circuits that emerges after repeated experience or observation, and demonstrate that these changes can generalize to novel and uncertain contexts. In ethological settings, this plasticity may have evolved to facilitate adaptive self-defensive strategies in situations in which aggression levels in nearby conspecifics are high, giving observers a leg up in promoting self-protection. However, these data may also shed light on the underlying circuit mechanism by which extreme levels of aggression exposure in humans may lead to long-lasting maladaptive behaviours. Methods Mice All animal procedures were approved by the Princeton University Institutional Animal Care and Use Committee and were in accordance with National Institutes of Health standards. For multifibre photometry recording studies, we used Vgat <sup +/− </sup ref. <sup 52 https://www.nature.com/articles/s41586-026-11217-9 ref-CR52 </sup mice backcrossed with CD1 Charles River, Crl:CD1 ICR , 022 male mice for a minimum of three generations, wild-type BALB/cJ males Taconic, 1-99 , and wild-type CD1 male mice. Vgat <sup +/− </sup and wild-type CD1 mice that were used as experimental subjects for freely moving or observational assays were sexually naive and aged between 12 and 18 weeks. BALB/cJ mice that were used as social targets for freely moving assays were sexually naive and aged between 8 and 26 weeks. Aggressor animals used during hard fights were single-housed aggression-experienced animals aged between 16 and 25 weeks. Mice aged between 15 and 16 weeks used for optogenetic stimulation for simulated fights were Esr1-cre +/ <sup − </sup ref. <sup 53 https://www.nature.com/articles/s41586-026-11217-9 ref-CR53 </sup male mice Jackson Laboratory, 017911 backcrossed to CD1 Charles River, Crl:CD1 ICR , 022 male mice for a minimum of three generations. Only male mice were used in this study. Mice were housed under a 12 h–12 h light–dark cycle with experiments taking place exclusively during the dark phase. Food and water were given ad libitum. The ambient temperature was maintained at 21–26 °C with a humidity of 30–70%. Before initiating fibre photometry experiments, Vgat <sup +/− </sup and wild-type CD1 mice used for aggression and/or observation experiments were socially isolated for at least 1 week. All animals were bred within the Princeton University animal facilities or purchased from Taconic Biosciences or Charles River. Animals were assigned to the EXP, OBS and NON groups randomly, constrained by the availability of animals of the appropriate genotype at the time of group assignment. Experimenters were blinded to group identity during behavioural scoring and histological verification. Behavioural assays Animals undergoing fibre photometry recordings were subjected to longitudinal resident–intruder experiments under varying conditions. For 8 consecutive days, animals undergoing longitudinal aggression experience EXP received unconstrained access to three different male BALB/cJ mice across three separate 9 min sessions, each preceded by a 1 min baseline without social stimuli. These interactions took place in the homecage of the experimental animals Optimice Auto-Water Cage Bottom with SE Lab Group cage, 34.3 cm length × 29.2 cm width × 15.5 cm height . For longitudinal observation OBS , a cage extender made of black acrylic and clear plexiglass was built to match the dimensions of the homecage, along with a compartment in the front of the cage extender designed to hold each observer mouse 10 cm length × 13 cm width × 17 cm height . The observer platformed compartment was separated from the cage floor through a clear plexiglass barrier with multiple horizontal cutouts allowing for visual and olfactory access of the cage bottom. Every OBS mouse was paired with a unique CD1 male mouse, with the latter undergoing 8 consecutive days of three back-to-back resident intruder sessions in its own homecage. For each day, each observer mouse was placed in the observer compartment, and the cage extender housing the compartment was placed on top of the paired animal’s homecage. Before initiating resident–intruder sessions, OBS mice and the partnered CD1 mice were allowed up to 15 min of habituation while separated by the plexiglass barrier. Next, each resident–intruder session consisted of 9 min of social interaction between the paired CD1 mouse and a different male BALB/cJ mouse, each preceded by a 1 min pre-intruder baseline. The control animals undergoing non-social experience NON were longitudinally exposed to an inanimate toy resembling a mouse over the same timespan as the EXP and OBS mice, and across the same number of sessions per day. Note that, while throughout the training period OBS individuals did not engage in social behaviour, photometry recordings were still performed in these individuals to mirror any bleaching-related signal decays present in EXP and OBS individuals after longitudinal imaging. Animals were assigned to the EXP, OBS and NON groups randomly. Blinding during data collection was not possible because group identity was evident from the experimental configuration. On day 9, each animal from each group was subjected to the hard fight, which consisted of three back-to-back 4–5 min resident–intruder sessions in their homecage, preceded by a 1 min solitary baseline. In each session, each animal received unconstrained social access to a unique undefeated CD1 male screened a priori for aggression against other CD1 males. Two EXP animals, inadvertently exposed to CD1 aggressors outside the scheduled training protocol, were excluded from hard fight behavioural analyses to avoid confounding protocol-driven effects with this uncontrolled exposure final behavioural, n = 10 ; neural data from these animals were retained for all neural analyses final neural, n = 12 , as the exclusion was motivated by behavioural interpretability rather than data quality. For comparison of familiar and novel males during observation, the mice were first defeated in their homecage by a unique CD1 male that had been previously screened for aggression against other CD1 males. This defeat session consisted of a 9 min social interaction preceded by a 1 min pre-intruder baseline. Afterwards, each defeated animal was placed in the observer compartment and was allowed to watch four consecutive resident–intruder sessions. Two of these resident–intruder sessions of the familiar CD1 aggressor while the other two were a novel CD1 male that the observer had not previously interacted with. All four resident–intruder sessions consisted of 9 min of social interaction between the novel or familiar CD1 mouse and a different male BALB/cJ mouse, each preceded by a 1-min pre-intruder baseline. The order of the novel and familiar sessions was randomized for each observer. Subsequent neural analyses were performed only for the first session of each type. For simulated attack observation, we unilaterally expressed ChR2 H134R -mCherry in the VMHvl of Esr1 <sup +/− </sup CD1 male mice. After viral expression, we screened for optogenetic stimulation evoked attack on an inanimate object toy mouse <sup 53 https://www.nature.com/articles/s41586-026-11217-9 ref-CR53 </sup . To assess whether the observers would respond differently to authentic attacks on a subordinate like the OBS group compared with a simulated attack on a toy mouse, the observers were presented with trials for authentic and stimulated aggression. The session structure consisted of a 1 min pre-stimulation baseline followed by 5 min during which optogenetic stimulation was delivered intermittently to evoke attack-like motor sequences toward the toy. Immediately after the toy session, the same observer animal then watched a standard resident–intruder session in which the same Esr1 <sup +/− </sup resident received a live male BALB/cJ intruder, preceded by a 1-min pre-intruder baseline. This within-animal design permitted direct comparison of observer neural responses during optogenetically evoked aggression toward an inanimate target versus naturalistic aggression toward a live conspecific, while controlling for the identity of the observed aggressor. Behaviour acquisition Social interactions were recorded using customized multi-camera systems, including top and side cameras 40 Hz, FLIR, Spinview . These were controlled with a data acquisition device DAQ and associated software Measurement Computing , which further coordinated the FP3002 CMOS camera in synchrony. The top camera was centred 50 cm above the homecage of experimental mice for each recording session, while the side camera was placed at a height of 30 cm and within 3 cm of the observer compartment. Black paper-chip bedding was used for every recording and miscellaneous items such as food pellets were cleared from homecages before recording. For behavioural analysis, we used a top-down camera, as this is reflective of what the observers viewed from their raised platform. Behavioural analysis Behavioural annotation For supervised classification of attack behaviours during the training period, the ground truth was determined by manual annotations of videos scored with BORIS<sup 54 https://www.nature.com/articles/s41586-026-11217-9 ref-CR54 </sup v.9.3.2 . Only attack and fighting behaviours were annotated, defined as rapid forward movements from the resident toward the intruder mouse, such as chases or lunges. The number of annotated frames included 29,805 attack frames out of 759,074 total frames across 49 videos. For hard fight sessions, behavioural annotation of specific behaviours was performed manually and carried out blind to the group identity of the behaving mice. Five behavioural categories were annotated. The first two categories captured aggressive behaviours initiated by the experimental subject: attacks delivered without resistance, defined as forward lunges or chases toward the CD1 aggressor that evoked no retaliatory response the CD1 fled or remained passive ; and attacks delivered with resistance, defined as attacks that evoked active defence from the CD1 aggressor in the form of counter-biting or physical pushback. The next two categories captured the analogous conditions from the perspective of receiving aggression: attacks received without resistance, in which the experimental subject was attacked by the CD1 aggressor and did not retaliate or push back; and attacks received with resistance, in which the experimental subject actively defended itself against the CD1’s attack. The fifth category captured fleeing behaviour by the experimental subject, defined as rapid locomotion away from the CD1 aggressor following an attack, approach or investigation. Pose tracking SLEAP software v.1.2.2 was used to predict resident and intruder positions across social interactions. To do this, 6,305 frames from 141 social interaction videos were labelled for tracking of resident and intruder body positions. These 6,305 frames were divided into training and test sets with an 80:20 split. The following points for both resident and intruder animals were labelled and subsequently tracked for all social interaction videos: nose, head, neck, left ear, right ear, trunk, left hindpaw, right hindpaw, left forepaw, right forepaw, tailbase, bottom segment of tail, middle segment of tail, upper segment of tail and tail tip. Training of the SLEAP model was performed over 93 iterations using a bottom-up approach with custom parameters initialized with Unet and batch-size of 4 . Feature definition To define the posture of resident and intruder mice, frame-to-frame pose data were converted to the following behavioural features: 1. 1 Intruder body length: the Euclidean distance between the head of the intruder and its tailbase. 2. 2 Resident body length: the same as above but for the resident. 3. 3 Intruder body orientation: the angle spanning the centroid, head and nose of the intruder. 4. 4 Resident body orientation: the same as above for the resident. 5. 5 Intersection over union of resident and intruder: the overlap between the resident and intruder bounding boxes by dividing the area of their intersection by the area of their union. 6. 6 Intruder head speed: the instantaneous distance in intruder head position every 4 frames or 100 ms . 7. 7 Resident head speed: the same as above but for the resident. 8. 8 Intruder head speed lagged at ±3 s: the sum across lags of intruder head speed, computed within a ±3 s lag window. 9. 9 Resident head speed lagged at ±3 s: the same as above but for the resident. 10. 10 Intruder centroid speed: the instantaneous distance in intruder centroid position every 4 frames or 100 ms . 11. 11 Resident centroid speed: the same as above but for the resident. 12. 12 Intruder centroid speed lagged at ±3 s: the sum across lags of intruder centroid speed, computed within a ±3 s lag window. 13. 13 Resident centroid speed lagged at ±3 s: the same as above but for the resident. 14. 14 Intruder to resident head-centroid angle: the angle spanning the intruder’s neck, intruder’s head and resident’s centroid. 15. 15 Resident to intruder head-centroid angle: the angle spanning the resident’s neck, resident’s head and intruder’s centroid. 16. 16 Intruder to resident head-to-tailbase distance: the Euclidean distance between the head of the intruder and the tailbase of the resident. 17. 17 Resident to intruder head-to-tailbase distance: the Euclidean distance between the head of the resident and the tailbase of the intruder. 18. 18 Between centroid distance: the Euclidean distance between the centroids of the resident and intruder. 19. 19 Between head distance: the Euclidean distance between the heads of the resident and intruder. Distance-related features were converted from pixels to centimetres using custom software based on our behaviour camera’s stereo parameters and pixel depth information. Feature preprocessing To standardize feature processing across all sessions, the mean and s.d. of every feature for every recorded session were pooled and subsequently averaged. Features for each session were processed as follows: 1 removal of outliers 3 s.d. away from the session mean; 2 smoothing through a Gaussian filter of 0.275 s; 3 linear interpolation of missing values; and 4 z -scoring using the mean and s.d. pooled from collating features across sessions. The collated mean and s.d. values for every feature were used for z -scoring in individual sessions to capture a wider variance in feature dynamics as opposed to unit variance, which may reflect data fluctuations that are not scalable across samples. Supervised behaviour classification To identify attack behaviours across training period videos, we trained a supervised multilayer perceptron MLP on SLEAP-derived feature data using the Keras library Keras and TensorFlow v.2.11.0 . Supervised classification was applied specifically to training period sessions, in which aggressive interactions are initiated exclusively by the resident experimental subject against a subordinate BALB/cJ intruder, yielding unambiguous one-sided attack bouts. Given the large volume of training period recordings across hundreds of videos, automated classification was necessary to make behavioural annotation tractable at scale. A sequential model was constructed consisting of 1 an input layer shaped after the behaviour feature set; 2 two hidden layers interleaved with batch normalization layers to normalize activations from the preceding layer; and 3 an output layer with a sigmoid activation function for binary classification. The input and hidden layers each consisted of 512 units with ReLU activation functions; the output layer consisted of a single unit. The target matrix was a binary indicator for hand-annotated attack labels pooled from 49 videos, totalling 29,805 attack frames versus 729,269 non-attack frames. The feature matrix included the 19 pose-derived features described above for each video frame, along with 19 additional lagged features capturing the temporal dynamics of the existing feature set, including the instantaneous distance between resident and intruder centroids. As non-attack frames far outnumbered attack frames, attack frames were oversampled through a synthetic oversampling technique for time series data. This involved generating synthetic samples for both attack labels and accompanying feature matrices by randomly sampling with replacement and combining these synthetic samples with the original data to balance each class distribution while preserving the underlying temporal structure. The resulting distribution totalled 699,464 attack frames versus 729,269 non-attack frames. This distribution was divided into training and validation sets with a 67:33 split, used for model development and internal performance assessment. To evaluate generalization to entirely new data, model performance was additionally assessed on a held-out test set drawn from videos outside the original annotation pool totalling 13,747 attack frames versus 433,936 non-attack frames . The performance on this held-out set yielded a balanced accuracy of 0.92; the confusion matrix is reported in Extended Data Fig. 1a https://www.nature.com/articles/s41586-026-11217-9 Fig6 . Subsequently, probability distributions for attacks were generated for every training period social interaction video. The probability threshold for classifying a frame as attack versus non-attack was set to 0.25; attack bouts were additionally required to span at least 200 ms of consecutive frames and to be separated from adjacent bouts by at least the same minimum duration. Supervised classification of behaviour during training period sessions was not performed for NON mice, as individuals in this group were not engaged in social interaction. Instead, behavioural labels for NON animals were generated through a yoked control design in which labels from a randomly assigned EXP or OBS individual were paired to each NON individual for the corresponding recorded day and session. All behavioural and neural analyses aligned to attacks during the hard fight were derived exclusively from hand-scored annotations rather than supervised model predictions. Hard fight sessions involve bidirectional aggression between the experimental subject and a CD1 aggressor, rendering one-sided supervised classification inappropriate. Hand annotations for hard fight sessions were performed blinded to group identity and segregated into the five behavioural categories described above attacks delivered not resisted, attacks delivered resisted, attacks received not resisted, attacks received resisted and experimental subject fleeing . Unsupervised behaviour embedding and clustering To comprehensively characterize social behaviour in an unbiased manner, we adapted previous unsupervised behaviour classification strategies involving the reduction of high-dimensional behaviour features to a low-dimensional embedding and the definition of behaviours as high-density clusters within this embedding. To achieve dense clusters, we embedded the 19 behaviour features described above into a two-dimensional embedding using UMAP. To generate this embedding, an importance sampling approach was taken. This approach increased the likelihood of generating a final version of the embedding featuring rare or nuanced behaviours. Importance sampling involved a total of two rounds of UMAP reduction. In the first round, 80,000 video frames were uniformly sampled from 504 social interaction videos. As aggression can often happen briefly and therefore be under-represented in the total pool of feature frames used for reduction, an additional 20,000 feature frames aligned to aggression labels according to our supervised attack classifier were randomly added to the frame pool. This total pool of 100,000 feature frames was next embedded into a two-component UMAP manifold umap from umap-learn, v.0.5.5, n neighbors=50, min dist=0.1 , which was further binned into a 200 × 200 histogram, smoothed with a 2D gaussian kernel of 3.5 s.d. and parcellated into 21 clusters with watershed skimage.segmentation.watershed . As UMAP is nonparametric, a separate model was implemented to learn the mapping from the original 19-dimensional feature space to the new 2D UMAP space. An MLP sklearn.neural network.MLPRegressor, hidden layer size of 1,000 × 500 × 100 × 50 units was used for this purpose, allowing the data from every video frame to be represented in UMAP space and therefore fall into any of the 21 clusters previously defined. To generate the final UMAP embedding, over 20,000 frames were sampled from the 21 clusters generated previously. Specifically, 2 random frames from each cluster and for each recorded session were pooled together for a total of 21,168 frames 2 frames × 21 clusters × 504 videos = 21,168 . From these sampled frames, the 19-dimensional feature set was embedded into 2D UMAP space. Next, the full set of video frames were mapped into this final UMAP manifold using another MLP. The resulting embedding was turned into a 2D histogram, smoothed with a Gaussian kernel of 2 s.d. and divided into 13 final clusters with watershed. All behaviour analyses involving behaviour clusters are performed using this final UMAP segmentation. Note that unsupervised classification of NON mouse behaviour during the training period was not performed as individuals in this group were not socially interacting; instead, cluster labels for individuals in this group were generated through a yoked control design whereby labels from a randomly assigned individual in the OBS or EXP groups were paired to each NON group individual for each recorded day and session. Behaviour transitions and relative similarity of unsupervised behaviour profiles The expression of UMAP clusters was quantified through the occupancy fraction of time in cluster relative to whole recording session or persistence mean number of consecutive frames comprising a segment of a behaviour cluster of each cluster, in addition to each cluster’s probability of transitioning to every other cluster on a frame-to-frame basis. To quantify the transition structure of behaviour clusters Fig. 1m https://www.nature.com/articles/s41586-026-11217-9 Fig1 , pairwise transition probabilities were calculated as the average number of times the first cluster transitions into the second cluster. For each session, a 13 × 13 matrix containing pairwise transition probabilities for all 13 clusters was constructed. A global transition matrix was then generated by averaging the transition matrices of each individual from each group from both training period and hard fight sessions. To evaluate above-chance significance in pairwise transitions, we implemented a surrogate significance test. We first calculated the empirical transition matrix ${T} {ij}\in {R}^{k\times k}$, where k is the number of states, from UMAP behaviour state cluster labels. T <sub ij </sub denotes the probability of transitioning from cluster i to cluster j . We then constructed a null transition matrix T <sup 0</sup that preserves the diagonal elements of T but replaces all off-diagonal entries with uniform probabilities, such that ${T} {{ij}}^{0}=\frac{ 1-{T} {{ii}} }{ k-1 }$ for i ≠ j . Using T <sup 0</sup and the empirically fitted state duration distributions, we generated W = 10<sup 4</sup surrogate sequences with 10<sup 5</sup samples each and re-estimated transition matrix $\hat{T}$ for each sequence, yielding a null distribution for each off-diagonal element. Let u <sub ij </sub be the mean of $\hat{T}$ across the N surrogates. The two-tailed empirical P value is calculated as ${p} {{ij}}=\frac{ 1+{M} {{ij}} }{ N+1 }$, where M <sub ij </sub is the number of surrogates with ${| \hat{T}} {{ij}}-{u} {{ij}}| \ge | {T} {{ij}}-{u} {{ij}}| $. Effect sizes are reported as ${D} {{ij}}=| {T} {{ij}}-{u} {{ij}}| $, the absolute deviation on the include-self transition scale. All 156 off-diagonal cells were tested, and P values were corrected for multiple comparisons at FDR = 0.05. Finally, the resulting global transition matrix was sorted through unsupervised hierarchical clustering scipy.cluster.hierarchy.linkage using the complete linkage method and Jensen–Shannon JS distance metric, with optimal leaf ordering to enhance the interpretability of the dendrogram. The resulting dendrogram was visualized with a colour threshold of 0.8 to distinguish clusters. Two complementary representations of this transition structure were used in downstream analyses. First, an aggregate transition weight was computed by pooling pairwise transitions between clusters found to individually differ in occupancy between groups relative to NON Fig. 1n https://www.nature.com/articles/s41586-026-11217-9 Fig1 clusters 2, 3, 4, 5, 6, 7, 8, 10, 11 and 13 , and summing these pairwise weights within each origin-category-to-target-category pairing according to the behaviour category assignments established in Fig. 1m https://www.nature.com/articles/s41586-026-11217-9 Fig1 and Extended Data Fig. 2 https://www.nature.com/articles/s41586-026-11217-9 Fig7 conflict, investigative, asocial , yielding a single aggregate transition weight for each of the nine possible category-to-category transitions Fig. 1o https://www.nature.com/articles/s41586-026-11217-9 Fig1 . Second, the full transition matrix all 156 off-diagonal pairwise transitions, without category aggregation or cluster filtering was row-normalized, such that each row summed to 1 and represented a valid probability distribution over target clusters conditional on the origin cluster; this row-normalized matrix was flattened to a single vector per individual per session and used as the transition-probability input to the JS divergence analysis described below. To assess whether behaviour occupancy, persistence or transition probabilities were predictive of group identity in the hard fight, a leave-one-session-out cross-validated approach sklearn.model selection.LeaveOneOut using support vector machine classifiers sklearn.svm.SVC was used. For each session, three kinds of vectors were generated: 1 a 13 × 1 vector in which each element represented the average occupancy each cluster; 2 a 13 × 1 vector in which each element represented the average persistence of each cluster; and 3 an 81 × 1 vector containing the flattened pairwise transition probabilities for 9 UMAP clusters identified as social behaviours. For each vector type above, a separate leave-one-out cross-validation procedure was implemented. Here a new classifier was iteratively constructed in each fold consisting of 68 vectors of occupancy, persistence or transition probabilities as the training set and a single held-out vector for the classifier to predict on. The balanced accuracy score of all held-out predictions was subsequently saved. Finally, to assess statistical significance of this leave-one-out cross-validation strategy, a null distribution was created by shuffling group identity labels 1,000 times and repeating the cross-validation process, recording the accuracy score for each shuffle. The P value for each model’s performance was computed as the proportion of shuffled accuracy scores greater or equal to the test held-out accuracy score. To assess whether OBS and XPO animals were behaviourally distinguishable from one another during the hard fight, the same leave-one-session-out SVM cross-validation framework described above was applied to decode OBS versus XPO group identity using cluster occupancy, persistence and pairwise transition probability vectors as separate feature sets. Statistical significance of the balanced accuracy score assessed against a permutation null distribution as described above. Furthermore, to further assess the degree to which OBS and XPO animals individually resembled EXP or NON behavioural profiles, a separate SVM classifier sklearn.svm.SVC, with probability=True was trained on EXP and NON animals using cluster occupancy, persistence or pairwise transition probabilities as feature sets. Occupancy, persistence and pairwise transition probability vectors were averaged across all three sessions of the hard fight to yield per-animal input sets. These classifiers were then used to predict on similar behaviour vectors for held-out OBS and XPO test points in which each individual from these groups was assigned a probability of being classified as EXP-like by the trained model. A Brunner–Munzel test scipy.stats.brunnermunzel was applied to the continuous EXP-classification probability scores to provide a test that uses the full distribution rather than a binary threshold. This analysis was performed separately for each of the three feature sets occupancy, persistence and transition probabilities . To calculate behavioural similarity between individuals in each group during hard fight sessions, the above vectors containing averaged occupancy and row-normalized transition probability variables were used. For every individual at a session level in the OBS group, similarity in occupancy or behaviour transitions to every individual at a session level in the EXP or NON groups was computed through the JS divergence of the occupancy or transition probability vectors of each pair of individuals. JS divergence here is employed as follows: where P represents the probability distribution corresponding to either behaviour occupancies or pairwise transition probabilities of an OBS individual, Q represents the same variable for an individual from either the EXP or NON groups, M is the pointwise arithmetic mean of P and Q that is, M = 1/2 P + Q , and D <sub KL</sub is the Kullback–Leibler divergence. The Kullback–Leibler divergence is defined as: JS divergence is disproportionately sensitive to high-probability elements of the distributions being compared, and behavioural cluster occupancy is unevenly distributed across clusters Fig. 1n https://www.nature.com/articles/s41586-026-11217-9 Fig1 , with asocial/avoidance clusters occurring at substantially higher occupancy than conflict and investigative clusters. To prevent this imbalance from disproportionately driving the occupancy-based JS divergence values, occupancy-based JS divergence Fig. 1q https://www.nature.com/articles/s41586-026-11217-9 Fig1 was computed separately within each of the three behaviour categories established in Fig. 1n https://www.nature.com/articles/s41586-026-11217-9 Fig1 conflict, investigative, asocial/avoidance and averaged across categories with equal weight. Furthermore, a similar strategy was adopted for transition-based JS divergence Fig. 1s https://www.nature.com/articles/s41586-026-11217-9 Fig1 . Pairwise transition probabilities are likewise unevenly distributed across cluster pairs Fig. 1m,o https://www.nature.com/articles/s41586-026-11217-9 Fig1 , with certain transitions occurring at substantially higher probability than others. To prevent this imbalance from disproportionately driving the transition-based JS divergence values, JS divergence was computed independently for each row of the transition matrix that is, separately for each of the 13 origin clusters and averaged across rows with equal weight. Finally, for all JS divergence comparisons occupancy and transitions , statistical tests were performed at the level of individual animals. For each individual, the category- or row-averaged JS divergence values computed against every individual, and across every session, in the comparison group were further averaged to yield a single value representing that individual’s mean divergence from the comparison group. It is these per-individual mean values, rather than the underlying session- or pair-level divergence values, that were used in the paired t -tests or Wilcoxon signed-rank tests described in the corresponding figure legends. Note that JS divergence is 0 when comparing equal distributions and larger values indicate larger divergence. Therefore, as presented in Fig. 1q,s https://www.nature.com/articles/s41586-026-11217-9 Fig1 , values closer to 0 reflect greater relative similarity between hard fight sessions. To visualize cross-cluster differences between groups Figs. 1l https://www.nature.com/articles/s41586-026-11217-9 Fig1 and 5n https://www.nature.com/articles/s41586-026-11217-9 Fig5 , raw per-cluster occupancy differences between groups per-animal, session-averaged were normalized by that cluster’s grand-mean occupancy across all groups compared, then scaled to a shared −1, 1 range by the largest such relative difference in the panel. This normalization affects visual presentation only; statistical comparisons between groups are reported separately. Supervised video-based classification of observer behaviour Model architecture We developed a deep neural network to classify observer behavioural states directly from raw video frames, consistent with previous machine learning approaches to behavioural classification<sup 55 https://www.nature.com/articles/s41586-026-11217-9 ref-CR55 , 56 https://www.nature.com/articles/s41586-026-11217-9 ref-CR56 </sup . The model processes short video clips 20 frames, 0.5 s at 40 fps and outputs probability scores across seven behaviour labels: attention, grooming, scratching, still, investigate, rearing and turning. The model architecture consists of two sequential components: a spatiotemporal encoder and a classifier Extended Data Fig. 5a https://www.nature.com/articles/s41586-026-11217-9 Fig10 , which is consistent with common machine learning solutions in animal behaviour<sup 55 https://www.nature.com/articles/s41586-026-11217-9 ref-CR55 , 56 https://www.nature.com/articles/s41586-026-11217-9 ref-CR56 </sup . The model is trained with Pytorch library v.2.5.1 . The encoder is a four-layer 3D convolutional neural network that transforms raw pixel values into high-level feature representation. Each layer applies 3D convolutions with kernel size 3 × 3 × 3, followed by batch normalization, ReLU activation, max pooling 2 × 2 × 2 and a spatiotemporal attention module described below . Channel dimensions double at each layer 64 → 128 → 256 → 512 , while spatial and temporal dimensions are reduced through pooling, progressively capturing higher-order representations of posture and movement. The final feature map is reduced to a 512-dimensional vector using global average pooling. The classifier is a three-layer feedforward network 512 → 256 → 128 → 7 with dropout regularization p = 0.3 between layers. Output logits are converted to class probabilities through softmax, and each video segment is assigned to the behaviour label with the highest probability. Spatiotemporal attention mechanism A key component of the encoder is a spatiotemporal module applied after each convolutional block, which is adapted from CBAM<sup 57 https://www.nature.com/articles/s41586-026-11217-9 ref-CR57 </sup for video data. The attention module allows the network to dynamically weight the importance of different spatial regions and specific frames within the data segment through separate stages explicitly. First, a channel attention step computes a global descriptor for each feature channel with adaptive average pooling, passes it through a bottleneck reduction ratio = 8 with a ReLU nonlinearity, and projects back to the original channel dimension. The resulting sigmoid-activated weights selectively amplify informative feature channels while suppressing less relevant ones. Second, a spatiotemporal attention step applies a 3D convolution across the channel-weighted feature map to produce an attention map, followed by sigmoid activation. This map assigns larger weights to behaviourally relevant regions for example, hind limbs during scratching, the head during grooming while suppressing background areas Extended Data Fig. 5b https://www.nature.com/articles/s41586-026-11217-9 Fig10 . Training procedure The model was trained end-to-end using the AdamW optimizer with different learning rates per component: the classifier was trained at a fixed rate of 3 × 10<sup −4</sup , while the encoder learning rate was warmed up from 7 × 10<sup −5</sup to 3 × 10<sup −4</sup over 20 epochs following an exponential schedule. This strategy allows the feature encoder to adapt gradually while the classifier learns to interpret its evolving representations. Models are trained for 300 epochs with an effective batch size of 256 using cross-entropy loss term with label smoothing α = 0.1 until validation loss plateaued. Gradient clipping max norm = 1.0 and mixed-precision training FP16 were used during training for stability and efficiency. All of the models fully converged, reaching approximately 86% validation accuracy and 92% training accuracy. The model snapshot with the best validation accuracy was used for inference and downstream analysis. To improve generalization and encourage the model to learn body-orientation-invariant representations, we applied stochastic augmentations in the spatial and temporal domains during training. Each video clip was randomly assigned to one of four augmentation modes: spatial only 30%; random rotation up to ±45°, random erasing of 2–33% of image area and random cropping , temporal only 30%; frame-order reversal , combined spatial and temporal 30% or no augmentation 10% . Validation and test samples were processed without augmentation. Dataset and evaluation We manually annotated observer behaviour across multiple video sessions with BORIS v.9.3.2 <sup 54 https://www.nature.com/articles/s41586-026-11217-9 ref-CR54 </sup , labelling each frame with one of seven behavioural labels: rearing upright posture with front limbs contacting the chamber walls ; turning fast reorientation of body heading within the chamber ; investigate active locomotion or head movements not directed toward the perforated wall ; still stationary idle posture not oriented toward the perforated wall ; grooming rhythmic front-limb movements directed at the head or ears ; scratching rhythmic hind-limb movements directed at the rear of the body ; and attention close proximity to the perforated wall with head oriented toward the barrier slits . The full dataset contains around 18,000 frames in total rearing, 6%; turning, 5%; investigate, 7%; still, 10%; grooming, 22%; scratching, 8%; attention, 43% . For model evaluation, we used fivefold cross-validation with 20% frames in validation and 80% frames in training set, stratified by labels. For main figures and downstream analyses, the seven classes are merged into four major categories: exploratory rearing, turning, investigate , idle still , self-directed behaviours grooming, scratching and attention attention . Model performance was assessed using the area under the receiver operating characteristic curve ROC-AUC , with metrics reported as the mean ± s.d. across five validation folds Extended Data Fig. 5c https://www.nature.com/articles/s41586-026-11217-9 Fig10 . Surgery Viral strategy At 12–18 weeks of age, Vgat <sup +/− </sup mice were anaesthetized isoflurane at 3–5% for induction and 1–3% for maintenance and placed into a stereotaxic frame before proceeding with injections and optic fibre implanting. A 1:1 300 nl cocktail of Cre-out GCaMP6f<sup 58 https://www.nature.com/articles/s41586-026-11217-9 ref-CR58 </sup Ef1a-Cre out-GCaMP6f and Cre-activated jRCaMP1b pAAV.Syn.Flex.NES-jRCaMP1b.WPRE.SV40 was delivered unilaterally into each of the following targeted brain regions: PrL anteroposterior AP −1.40 mm, mediolateral ML 0.35 mm, dorsoventral DV −1.70 mm , vLS AP −2.60 mm, ML 0.60 mm, DV −3.25 mm , mPOA AP −3 mm, ML −0.40 mm, DV −4.75 mm , BNST AP −3.30 mm, ML 0.85 mm, DV −3.75 mm , AH AP −3.7 mm, ML −0.50 mm, DV −4.80 mm , VMHvl AP −4.65 mm, ML 0.72 mm, DV −5.45 mm , LHb AP −4.70 mm, ML −0.33 mm, DV −2.65 mm , MeA AP −4.70 mm, ML −2.25 mm, DV −4.90 mm , PA AP −5.50 mm, ML −2.40 mm, DV −4.95 mm , PMv AP −5.50 mm, ML −0.58 mm, DV −5.55 mm , lPAG AP −7.80 mm, ML 0.50 mm, DV −1.65 mm . An additional infusion of 150 nl of GRAB DA1m hSyn-GRAB DA1m-WPRE-hGHpA was delivered to the NAc AP −1.80 mm, ML 1.15 mm, DV −3.70 mm . Surgery coordinates for each area were calculated relative to the skull surface in the posterior edge of the rostral rhinal vein. All regions included in the final analyses had correct fibre placement based on histology. For chemogenetic experiments, Vgat <sup +/− </sup or wild-type Vgat <sup −/− </sup mice received bilateral infusions of 150 nl of an AAV2 vector encoding a Cre-dependent excitatory DREADD Ef1a-DIO-hM3Dq-mCherry; Addgene, 50460; PNI Viral Neuroengineering Lab into the MeA, vLS, PMv and AH. This strategy selectively expressed the G<sub q</sub -coupled hM3Dq receptor in Vgat<sup +</sup inhibitory neurons in Vgat <sup +/− </sup mice, while serving as a control in wild-type animals lacking Cre recombinase activity. For optogenetic experiments, Esr1 <sup +/− </sup mice received a unilateral infusion of 150 nl of an AAV2/5 vector encoding a Cre-dependent channelrhodopsin Ef1a-DIO-hChR2 H134R -mCherry-WPRE-hGHpA; PNI Viral Neuroengineering Lab into the right VMHvl, enabling selective photoactivation of ESR1<sup +</sup neurons in this region. Multi-site fibre strategy After viral infusions, a custom-built multi-fibre implant<sup 59 https://www.nature.com/articles/s41586-026-11217-9 ref-CR59 </sup containing twelve 200-μm-core optic fibres 220 μm cladding, 0.37 NA, THORLabs was fixed to the skull with Metabond Parkell . This implant was lowered into the brain to a depth of 5.55 mm relative to the brain surface of the VMHvl using a custom stereotaxic attachment designed to hold multi-fibre implants. Multi-fibre implants are 3D-printed ABS-Like Micro Resolution, Protolabs and designed with fibre through-holes in AP and ML locations matching those used for stereotactic viral infusions. Before fixing the fibres in the implant, these are first cleaved with a manual fibre cleaver THORlabs and subsequently cut with a ruby scribe to match the specific length of each area’s depth in the brain, with added length accounting for the widths of the implant and each area’s craniotomy. A summary of all coordinates and measurements in mm used for viral deliveries and implant construction is provided in Supplementary Table 2 https://www.nature.com/articles/s41586-026-11217-9 MOESM1 . Before implantation, bulk light efficiency for every implant was measured, with implants with an average efficiency below 70% getting discarded. After surgery completion, the mice were allowed to recover for at least 3 weeks before the start of recording experiments. Multi-site fibre photometry Recording To record from up to 12 sites simultaneously, we used a custom 12-fibre bundled, branching patchcord 200 μm fibre, 220 μm cladding, 0.37 NA, Doric . One end of the patchcord terminated in an SMA connector in a honeycomb pattern, with the other end affixing all branching fibres in a custom 3D-printed connector at coordinates matching the fibre placements in the multi-fibre implant. The implant fibres were aligned to the patchcord fibres in the connector using four 1 mm steel dowel pins McMaster , and the two components affixed through 3–4 screws Open Ephys ShuttleDrive . Recordings were made through a custom written data acquisition program in DAQAmi Measurement Computing and using the Neurophotometrics FP3002 system in combination with Bonsai. LEDs delivering three excitation wavelengths 470 nm for optical excitation of GCaMP6f and DA1m GRAB, 560 nm for optical excitation of jRCaMP1b, and 415 nm for a calcium- and dopamine-independent control, light intensity of around 70 μW for each fibre tip of the implant were interleaved at 40 Hz for a total 120 Hz sampling rate throughout recordings. Fluorescence emission was split at 552 nm into green and red channels and focused onto a CMOS sensor for detection. Photometry signals in arbitrary units were collected from 12 sites drawn separately in the green GCaMP and red RCaMP channels across the imaged honeycomb pattern of fibres. Signal processing The mean fluorescence in arbitrary units was collected as raw data for every session. To normalize this signal and correct for potential motion artifacts, we performed processing on individual photometry signals as follows: 1 baseline subtraction using an adaptive iteratively reweighted penalized least squares algorithm<sup 60 https://www.nature.com/articles/s41586-026-11217-9 ref-CR60 </sup ; 2 z -score standardization; and 3 signal subtraction using motion-artifact control signals from 415 nm excitation. For step 3, specifically, we fitted a multilinear regression sklearn.linear model.Ridge for each individual photometry signal using as predictors all calcium-insensitive control signals from 415 nm optical excitation taken from the same recorded session<sup 59 https://www.nature.com/articles/s41586-026-11217-9 ref-CR59 </sup . The regularization strength alpha for each fitted model was determined individually for each session through a grid search. The predicted control signal was then used to subtract the real photometry signal. The resulting corrected signals from this process were finally normalized by their mean value during a 1 min baseline period before any social interaction, yielding the z- scored Δ F / F signal. To remove residual high-frequency artifacts, z- scored Δ F / F traces were subsequently smoothed using a Savitzky–Golay filter scipy.signal.savgol filter; window length = 7 samples, polynomial order = 3 . As SLEAP-derived behavioural features see the ‘Pose tracking’ section above were acquired at a lower temporal resolution than the photometry system, smoothed neural traces were then resampled to match the temporal dimension of the corresponding behavioural feature series for each session: each channel was linearly interpolated onto a new, evenly spaced set of timepoints spanning the same session duration as the original recording, with the number of resampled timepoints set to match the number of frames in the corresponding SLEAP-derived feature series. All analyses were completed on the resulting resampled z- scored Δ F / F data. Optogenetic stimulation To optogenetically activate Esr1 <sup +/− </sup neurons in the VMHvl, a blue laser 460 nm, 7.0–7.2 mW was connected to a 200 μm diameter patch cord that was fastened to the mouse’s implanted fibres through ceramic sleeves. For all stimulation, we delivered 20 ms pulses given at 20 Hz 40% duty cycle . Laser activation was controlled through an Arduino microcontroller system that received TTL trigger signals from manual button presses, enabling experimenter-initiated optogenetic stimulation during behavioural sessions. The Arduino was programmed to convert each button press into a train of TTL pulses with the specified temporal parameters 20 ms on, 30 ms off , which directly gated the laser output. Neural analysis Time-shifting of attack signals To identify a population-wide single timepoint for observed attack alignment on a trial-by-trial basis, a shift-only time-warping approach<sup 38 https://www.nature.com/articles/s41586-026-11217-9 ref-CR38 </sup was used Extended Data Fig. 6 https://www.nature.com/articles/s41586-026-11217-9 Fig11 . This approach addresses temporal uncertainty arising from variability in the timing of neural responses relative to classifier-defined attack onsets. Moreover, moment-to-moment fluctuations in attentional state introduce further latency variability in neural engagement with each observed attack. Traditional PETHs aligned to attack onset therefore average across this temporal jitter, potentially obscuring consistent neural dynamics that are present but variably shifted across trials. Time-shifting was implemented using the ShiftWarping algorithm from the affinewarp package v.0.20.0 , extended through a custom RobustTimeWarping subclass that adds artifact trial detection, boundary shift correction and template-based alignment of held-out animals. The algorithm learns a single scalar time shift per trial that translates the neural signal forward or backward in time, without stretching or compressing temporal structure. Optimization proceeds iteratively: beginning from the trial-mean as an initial template, the model alternates between 1 finding the optimal per-trial shift that minimizes reconstruction error relative to the current template; and 2 updating the template as the mean of all shift-corrected trials. This loop is repeated for a fixed number of iterations specified per group below until convergence. Critically, a single time shift is determined jointly from the activity of all simultaneously recorded neural populations; brain regions are never shifted independently. Before model fitting, trials exhibiting mean z -scored activity exceeding a specified threshold during an early artifact window frames 0–40, corresponding to the pre-attack baseline period were excluded. This step removes trials where residual activity from immediately preceding attacks contaminates the baseline, which would bias the learned template toward elevated pre-event activity. For the OBS group, a threshold of 4 z -score units was applied; for the EXP group, this threshold was set permissively high effectively disabled given the lower rate of such artifacts in directly experienced aggression. EXP group: attack-aligned neural signals −6 s to +6 s, window size = 480 frames at 40 Hz were extracted for all animals in the EXP group across all training period sessions. Trials containing any NaN values due to missing fibre placements were removed before model fitting. A RobustTimeWarping model was fit to the resulting clean tensor trials × time × 23 populations with a maximum allowable shift of ±0.05 × window duration that is, ±0.5 s; maxlag = 0.05 , no smoothness or L2 regularization, and a single optimization iteration. For animals with one or more missing brain regions due to incorrect fibre placement, the learned template from the full-cohort model was used to infer optimal per-trial shifts through a separate align to template procedure. In this procedure, missing regions were temporarily filled with data from the nearest available region to enable shift optimization, then restored as NaN in the final time-shifted output. OBS group: aRobustTimeWarping model was fit to clean OBS trials NaN-free from seven of the nine observer animals, with the remaining two animals collected after the primary model was fit. The model was fit with maxlag = 0.18 ±2 s maximum shift , smoothness regularization scale = 0.1, warp regularization scale = 0.1, L2 regularization scale = 0.1, artifact threshold = 4 z -score units over frames 0–40 and 5 optimization iterations. After model fitting, the learned template was used to align the two held-out animals’ clean trials by align to template. Animals with one or more missing brain regions due to incorrect fibre placement were handled using the same nearest-available-region substitution procedure described for the EXP group, with missing regions restored as NaN in the final time-shifted output. Time-shifted data from the primary fit and held-out animals were concatenated into a single combined OBS tensor for all downstream analyses. Observer attention state labels ATTN+/ ATTN− were propagated through the same trial indexing pipeline to enable subsequent attention-filtered analyses. NON group: NON animal neural signals, aligned to timepoints derived from OBS attack labels yoked design , were time-shifted by applying the same fitted OBS model template via fit transform with identical hyperparameters. This ensures that any temporal structure revealed in OBS signals under the same model reflects genuine signal structure rather than an artifact of the alignment procedure. For neural signals recorded during attacks toward a toy versus a live conspecific, and during familiar versus novel conspecific encounters, the OBS training period model was reset and re-fit to the primary OBS dataset, and the resulting template was used to align these held-out datasets through align to template. This ensures that temporal alignment is governed by the same aggression-specific template derived from observer neural dynamics rather than from the held-out behavioural contexts themselves. Validation controls: to confirm that the learned time shifts reflect aggression-specific neural dynamics rather than generic movement statistics or arbitrary temporal structure in attack bouts, three control conditions were implemented. In the first synthetic control, attack labels were replaced with synthetic sequences generated by a Markov chain order = 10 trained on empirical attack data from 20 randomly selected sessions, preserving the statistical structure of attack timing transition probabilities and bout durations while destroying the temporal correspondence between labels and neural activity. The model learned state transition probabilities from the observed attack label sequences and additionally modelled the empirical distribution of attack bout durations, sampling bout lengths from this distribution during sequence generation to preserve realistic bout structure. A target attack rate of 0.1 was applied to match the approximate overall attack frequency observed in real sessions. To ensure a fair comparison between real and synthetic attack-aligned signals, synthetic trial counts were matched to real trial counts on a per-animal, per-day basis by randomly subsampling synthetic trials where the number of synthetic trials exceeded the number of real trials; sessions in which synthetic trials were fewer than real trials retained all available synthetic trials. Neural signals from OBS animals were then time-shifted using these synthetic labels under the same model. In the second speed control, attack labels were replaced with binary indicators of high resident centroid velocity threshold: 2 z -score units , with any overlap with true attack periods removed. The same OBS model was applied to signals aligned to these speed events. In the third yoked control, the OBS model was applied to NON animal neural signals using attack time indices derived from real OBS sessions, testing whether the model produces spurious structure when applied to neural data from animals that never observed aggression. In all three controls, the resulting time-shifted signals and their aligned PETHs were compared against those from real OBS attack trials. PETH computation: after time-shifting, PETHs were computed for each neural population and experimental group by averaging time-shifted signal tensors across trials, either per-animal yielding a time × populations × animals array or across all animals yielding a time × populations array pending the type of analysis. Baseline correction was applied by subtracting the per-trial or per-animal median activity during a pre-event window −5 s to −3 s before attack onset from each timepoint. For analyses of ATTN+ versus ATTN− trials, time-shifted OBS trials were filtered by requiring a minimum of 5 s of ATTN+ occupancy within the attack window 200 frames out of 480 at 40 Hz before averaging. Hard fight: conflict-aligned neural signals during the hard fight were time-shifted using the same RobustTimeWarping procedure described above, with a more conservative parameterization than that used for training-period alignment center shift = True, maxlag = 0.05, that is, a maximum allowable shift of ±0.05 × window duration . This approach was adopted for two reasons: first, hard-fight conflict bouts encompass a mixed composition of behaviour types attacks delivered and received, resisted and not resisted; Fig. 1d https://www.nature.com/articles/s41586-026-11217-9 Fig1 , and a tighter, centred shift constraint helps ensure consistent alignment across this heterogeneous trial composition rather than allowing the model to drift toward any one behaviour type; second, the hard fight yields substantially fewer trials per animal than the training period, and a less permissive shift window reduces the risk of overfitting the learned template to a sparse trial set. Missing-region handling, held-out-animal alignment by align to template. Decoding observer behaviour from population neural activity To test whether population neural activity in OBS animals carries information about the observer’s behaviour, we trained logistic regression classifiers on summary features extracted from windowed neural activity and evaluated their ability to classify behavioural categories. Neural recordings were segmented into 1.5 s windows 60 frames with a 0.75 s stride 50% overlap , and only windows in which a single behavioural category occupied ≥90% of frames were retained and labelled with that category. Within each window, we extracted 10 summary statistics from the activity of each neural population maximum, 90th percentile, 10th percentile, median, 10% trimmed mean, s.d., mean squared amplitude, s.d. and mean absolute value of the first-order difference, and s.d. of the second-order difference , yielding a 190-dimensional feature vector per window across the 19 recorded populations. Decoding was performed separately for each animal and behavioural category as a binary one-versus-rest classification. For each animal and category, a logistic regression classifier with L2 regularization C = 1 and balanced class weights was fit to features standardized using training-fold statistics. We evaluated model performance with stratified group fivefold cross-validation, in which windows were assigned to folds in blocks of 50 consecutive windows so that temporally adjacent, overlapping windows were not split between training and test sets. Cross-validation was repeated 10 times with different fold assignments, and the positive-class F1 score on held-out windows was averaged across folds and repeats to obtain one score per animal and category. We compared two decoder designs: an all-regions decoder trained on the features of all 19 populations, and a best single-region decoder. For the latter, a separate single-region decoder was trained on the features of each of the 19 populations, and for each animal and category the population whose decoder achieved the highest F1 score averaged across the 10 cross-validation repeats was selected; the F1 score of that decoder was reported. Chance performance was estimated by repeating the all-regions decoding procedure on randomly permuted labels 100 permutations per animal and category and averaging the resulting F1 scores. Classifiers, feature standardization, cross-validation and F1 scoring were implemented in scikit-learn. To identify the populations that contributed most to the all-regions decoder, we computed SHAP values<sup 61 https://www.nature.com/articles/s41586-026-11217-9 ref-CR61 </sup using the linear explainer from the SHAP package. For each animal and category, the decoder was refitted within the block cross-validation scheme described above, and interventional SHAP values were computed on held-out windows using the training fold as the background distribution. The importance of each population was quantified as the mean absolute SHAP value across held-out windows and that population’s 10 features, averaged across folds and repeats. Decoding condition and identity from single-population neural activity To assess whether attack-aligned neural activity in individual SBN populations carries information about the context, social target or identity of the observed aggressor, we trained logistic regression classifiers LogisticRegression from scikit-learn on single-region mean z- scored Δ F / F activity and evaluated their ability to decode binary condition labels on a trial-by-trial basis. For each decoding analysis, the feature for each trial consisted of the mean z- scored Δ F / F of a single neural population computed within an active attack window after subtraction of a pre-attack baseline median. Classifiers were trained and evaluated separately for each recorded region, and all analyses were restricted to regions that showed significant attack-aligned activity in OBS animals identified through the PETH comparison framework described above . Given the sparsity of trials per animal in certain experimental conditions, decoding was performed across all trials pooled from all animals rather than on a per-animal basis. For each region, a logistic regression classifier with balanced class weights was fit using a pipeline that included standard feature scaling. Regularization strength and penalty type L1 or L2 were selected by inner grid search using stratified k -fold cross-validation sklearn.model selection.StratifiedKFold . The number of folds for the outer cross-validation loop was determined empirically by evaluating classifier performance across a range of fold counts and selecting the value that maximized mean cross-validated accuracy, subject to a minimum of nine samples per fold. The final classifier was re-evaluated using the selected fold structure, and per-fold balanced accuracy scores were stored for subsequent significance testing. To determine whether decoding accuracy was significantly above chance, a one-sample t -test or Wilcoxon signed-rank test was applied to the distribution of per-fold scores against a chance balanced accuracy level of 0.5 one-sided, testing whether mean accuracy 0.5 . The following decoding analyses were performed: Authenticity of aggression conspecific versus toy target : observer animals watched either naturalistic aggression toward a live BALB/cJ intruder or optogenetically evoked attack-like behaviour directed toward an inanimate toy by an Esr1-cre animal expressing ChR2 in the VMHvl. Neural signals from both conditions were time-shifted using the OBS training period model before feature extraction, applying the same template-based alignment procedure described above. Familiarity with aggressor familiar versus novel conspecific : observer animals that had been previously defeated by a CD1 aggressor subsequently watched that familiar aggressor or a novel CD1 aggressor attacking a BALB/cJ intruder. Neural signals were time-shifted using the OBS training period model before feature extraction. Attack vigour—resident attack speed: observer neural responses were sorted trial-by-trial by the z -scored peak centroid speed of the resident aggressor during the attack window. Trials falling at or below the 10th percentile of resident attack speed were designated as low-vigour trials, and trials at or above the 90th percentile were designated as high-vigour trials. Attack effectiveness—intruder flight speed: the same trial-sorting and classification procedure was applied using the z -scored peak centroid speed of the intruder following each attack, as an index of attack effectiveness. Trials at or below the 10th percentile of intruder flight speed were designated as low-effectiveness trials, and trials at or above the 90th percentile were designated as high-effectiveness trials. Response decay To characterize the temporal dynamics of attack-aligned calcium signals across recorded neural populations, several decay metrics were computed on a per-animal basis from baseline-corrected mean PETHs. For each animal and each recorded population, the mean z- scored Δ F / F signal was computed by averaging across all time-shifted, attack-aligned trials within the training period, yielding a single mean PETH trace per animal per region. All of the subsequent decay metrics were derived from this trial-averaged trace. The peak of each PETH was identified within a search window spanning timepoints −0.5 s to +0.5 s around attack onset. To improve robustness against single-frame transients, a 3-frame 0.075 s uniform smoothing filter was applied to the trace before identifying the peak location; however, the actual peak value used for all metric calculations was taken from the unsmoothed trace at the identified peak frame. Traces of which the detected peak fell below the noise floor threshold were excluded from all decay metric computations for that region and animal, and the corresponding metric values were set to NaN. For each animal and region with a valid peak, t <sub 1/2</sub was defined as the time elapsed from the peak frame until the signal first crossed 50% of its peak value in the post-peak portion of the trace. To achieve a sub-frame temporal precision, linear interpolation was applied between the last frame above and first frame below the half-target value. If the signal never reached 50% of its peak value within the post-peak analysis window frames 0 to end of trace from peak , the trial was flagged as non-decaying and the t <sub 1/2</sub value was set to NaN rather than assigned the maximum window duration, to avoid inflating decay time estimates. Relative similarity of neural activity profiles To compare the similarity of population-level neural representations across behavioural states between groups, the mean z -scored Δ F / F of each recorded neural population was computed separately for each of the 13 UMAP behavioural clusters, yielding a 13 × 1 activity vector per population per session. Cluster membership at each timepoint was determined by the unsupervised behavioural labels assigned during UMAP segmentation. If no timepoints were assigned to a given cluster within a session, the corresponding vector element was set to zero. Note that activity vectors for OBS mice during the training period were drawn from neural time series filtered for positive attention states see the ‘Supervised video-based classification of observer behaviour’ section . Pairwise similarity between sessions from different individuals was quantified using cosine distance, defined as: where A and B represent mean activity vectors for two different individuals. Cosine distance measures divergence in the orientation of activity vectors rather than their magnitude, making it well-suited for identifying populations of which the behavioural representations share a similar structure regardless of overall signal scale. A cosine distance of 0 indicates identical vector orientation and therefore maximum similarity, while larger values indicate greater divergence. For training period sessions, activity vectors were mean-centred before computing cosine distance. Specifically, for each session and each neural population, the per-cluster activity vector was centred by subtracting its mean across clusters. Although cosine distance is intended to be invariant to vector magnitude, it is sensitive to differences in the origin of the activity vectors—that is, if two populations share the same representational structure but differ in their absolute baseline level, the vectors will point in different directions in high-dimensional space, inflating the apparent distance between them. Such baseline differences are expected when comparing activity vectors across contexts in animals performing radically distinct behaviours. Furthermore, these differing baseline conditions can shift the absolute level of neural activity, giving the activity vectors different effective origin points. Mean-centring removes this global offset, ensuring that cosine distance reflects differences in the shape of the representation across behavioural states rather than differences in baseline. All 13 UMAP clusters were included in the activity vectors for training period analyses. To further account for the temporal variability in neural responses during the training period—as evidenced by the trial-by-trial time shifts required to align attack-period neural activity see the ‘Time-shifting of attack signals’ section —pairwise distances during the training period were computed using a soft cosine distance measure rather than standard cosine distance. Soft cosine distance modifies the standard cosine distance by incorporating a cluster similarity matrix S that encodes the expected functional relationships between behavioural clusters: The matrix S was constructed by grouping clusters into three functional categories—aggressive clusters 1–4 , investigative clusters 5 and 7–9 and asocial clusters 6 and 10–13 —and assigning within-group similarity of 0.5 and between-group similarity of 0.0, with a diagonal of 1.0. This formulation reduces the penalty assigned to activity vector differences that fall within the same functional cluster group, reflecting the expectation that temporal jitter in neural responses may displace activity across closely related behavioural states without representing a true representational difference. As the temporal alignment of neural activity is more reliable during hard fight sessions where all animals are engaged in equivalent social interactions and the behavioural context is more constrained standard cosine distance was used for hard fight comparisons and mean-centring was not applied. Pairwise distance computation For each session, neural population and pair of animals, cosine distance or soft cosine distance, for the training period was computed between the corresponding 13 × 1 activity vectors. Distances were stored in a sessions × populations × OBS animals × other animals array. For statistical comparisons, this matrix was averaged jointly across the sessions and other-animals dimensions to yield a single cosine distance value per population per OBS animal, separately for the OBS-EXP and OBS-NON comparisons. After this reduction, each OBS animal contributed exactly one cosine-distance value per population for its comparison to EXP and one for its comparison to NON, constituting a paired sample both values are derived from the same animal . Network-level decoding of group identity from social behaviour-aligned neural activity To assess whether the multi-region pattern of neural activity during social behaviour distinguishes EXP from NON animals and to determine where OBS animals fall along this axis, a LDA classification approach was applied to population-level activity vectors aligned to social behavioural clusters during the hard fight, implemented using scikit-learn. For each animal and each session, the mean z- scored Δ F / F was computed for each recorded neural population across all frames assigned to social behavioural clusters conflict clusters 1–4 and investigative clusters 5 and 7–9 , yielding a per-session, per-population scalar value. A social engagement index was additionally computed for each session as the difference between mean activity during social clusters 1–5, 7–9 and mean activity during asocial clusters 10–12 , providing a single value per population per session that captures the relative modulation of each population by social versus non-social behavioural contexts. These values were averaged across sessions to yield one feature vector per animal, with each element corresponding to a single recorded neural population. Feature matrices were constructed separately for the full population set, excitatory-only and inhibitory-only subsets. An LDA classifier with automatic shrinkage regularization solver = ‘eigen’, shrinkage = ‘auto’ was fit on EXP label = 1 and NON label = 0 animals using each of the three feature subsets. Classifier performance was evaluated using leave-one-out LOO cross-validation, with balanced accuracy as the scoring metric. Statistical significance of LOO accuracy was assessed by permutation testing 1,000 permutations of group labels , with the observed LOO accuracy compared against the resulting null distribution. OBS animals were treated as held-out test points and were never included in classifier training; their feature vectors were projected onto the learned discriminant axis using the same imputer and scaler fit on the EXP/NON training data. To provide an interpretable measure of OBS placement relative to the trained groups, a normalized OBS position was computed as: where $\hat{p}$ denotes the mean projection for each group, such that a value of 0 indicates OBS animals are indistinguishable from EXP mice and a value of 1 indicates that they are indistinguishable from NON mice. Furthermore, to identify which inhibitory populations most strongly drove OBS placement on the EXP–NON axis, a leave-one-population-out procedure was applied to the inhibitory-only subspace. For each iteration, one inhibitory region was dropped from the feature set and the LDA was refit on the remaining regions. The shift in normalized OBS position ΔOBS position relative to the full inhibitory model was recorded for each dropped region, with positive values indicating that removing that region shifted OBS animals toward the NON end of the axis. To provide a complementary dimensionality-reduction view, PCA was applied to the same scaled EXP and NON feature matrices using the full population set. The first two principal components were retained, and OBS mice were projected as held-out test points onto the learned PC space. Decomposition of shared and unique variance in VMHvl activity across behavioural states To characterize the unique and shared contributions of each recorded SBN region to VMHvl excitatory population activity, we used an automatic relevance determination ARD regression framework with permutation-based variance decomposition<sup 41 https://www.nature.com/articles/s41586-026-11217-9 ref-CR41 , 62 https://www.nature.com/articles/s41586-026-11217-9 ref-CR62 , 63 https://www.nature.com/articles/s41586-026-11217-9 ref-CR63 </sup , applied separately for each behavioural cluster, session and animal. For each animal, neural activity across all sessions was concatenated into a trial × region × epoch array using a standardized formatting pipeline. A fixed set of eight predictor regions was defined a priori indices corresponding to VMHvl inhibitory, lPAG excitatory, PMv excitatory, PMv inhibitory, AH inhibitory, MeA inhibitory, BNST inhibitory and vLS inhibitory populations , chosen to span the recorded SBN while excluding the target region VMHvl excitatory . The target variable y was the VMHvl excitatory population activity vector for a given epoch, and the predictor matrix X was the corresponding activity matrix of the nine predictor regions. Epochs or clusters in which any predictor region contained exclusively NaN values, or in which fewer than 80 timepoints were available within a given behavioural cluster, were excluded from analysis. For activity corresponding to OBS mice in the training period, neural activity for each session was filtered for positive attention states a priori. For each behavioural cluster c within each epoch e , an ARD regression model was fit using threefold cross-validation KFold, shuffled . ARD regression was selected over standard ridge regression because it places independent Gaussian priors over each regression coefficient, enabling automatic pruning of irrelevant predictors while providing stable coefficient estimates under multicollinearity. Predictions from held-out test folds were concatenated to produce a full out-of-sample predicted time series, and the out-of-sample R <sup 2</sup was computed against the true target. Negative R <sup 2</sup values were clipped to zero. The final cross-validated R <sup 2</sup was stored as the full-model performance score for that cluster and epoch. Regression coefficients were averaged across the three folds to produce a single weight vector per predictor region. Unique variance permutation-based knockout : for each predictor region, unique variance was estimated by comparing full-model performance against a knockout model in which the values of that predictor were randomly permuted within the test set at prediction time, leaving the trained model weights intact. This procedure was repeated across 100 permutation iterations per predictor, and the mean knockout R <sup 2</sup was computed across iterations. The unique variance attributable to a given region was defined as the difference between full-model R <sup 2</sup and mean knockout R <sup 2</sup , clipped at zero: Marginal variance and shared variance: to estimate the marginal contribution of each predictor region in isolation, a separate ordinary least squares regression model was fit using only that single predictor, again under threefold cross-validation. Ordinary least squares was used in place of ARD for marginal models given that single-feature models preclude multicollinearity. The marginal R <sup 2</sup was computed from the resulting out-of-sample predictions, clipped at zero. Shared variance for each predictor was then defined as the difference between its marginal R <sup 2</sup and its unique variance: This quantity captures the portion of variance that a given region shares with the remaining predictors—that is, variance that can be explained by that region alone but is not uniquely attributable to it when the full predictor set is considered. To test whether the unique variance Δ R <sup 2</sup attributable to each predictor region changed systematically over the course of training, per-session unique variance values were modelled as a function of training day using a LME model, implemented in statsmodels mixedlm . For each recorded region and each experimental group EXP, OBS, NON separately, a model of the form value ~day + 1|animal was fit, with the day of training entered as a fixed effect and the animal identity as a random intercept, to account for repeated measurements within animals across the 8-day training period. The day coefficient from each converged model was tested against zero using a Wald test. For excitatory predictors, for which no a priori directional hypothesis was specified, this test was two-sided. For inhibitory predictors, consistent with the a priori hypothesis that inhibitory unique variance declines over the course of training, a one-sided test was used: the two-sided P value was halved when the estimated slope was in the hypothesized negative direction, and set to 1 − half the two-sided P value otherwise, preventing an effect in the direction opposite the hypothesis from being reported as spuriously significant. Behaviour feature encoding To characterize how the behavioural features were encoded in the neural activities across the network, we fit independently for each feature and each neural signal a linear Gaussian generalized linear model GLM with a smoothing regularizer that is chosen based on cross-validation. The input to the model is a one-hot vector representation of the feature value, which was binned in 20 total bins. The fitted GLM weights reflect the tuning curve for that specific region and feature, with positive weights suggesting an increase in activity and negative weights a decrease. To test for group differences in tuning, we used a leave-one-animal-out procedure: for each held-out animal, group-specific tuning curves were fit to the remaining animals, and the cross-validated R <sup 2</sup of the held-out animal’s neural activity was computed using the tuning curve of its own group. For each group, feature and region, a one-sided Wilcoxon signed-rank test was used to determine whether R <sup 2</sup was significantly greater than P < 0.05 . Only feature–region pairs for which at least two of the three groups EXP, OBS, NON showed R <sup 2</sup significantly above 0 were analysed further. For these pairs, per-group tuning curves were averaged over the lowest five and highest five bins of the feature, and each pairwise group comparison was performed using a two-sided independent-samples t -test to assess differences in tuning at the extremes of the feature range. Whole-brain histology for photometry fibre recovery and viral expression assessment After the completion of our photometry and behavioural studies, animals were deeply anaesthetized by injection of a ketamine and xylazine cocktail, and perfused with 4% paraformaldehyde PFA dissolved in 1× PBS. Brains were extracted and post-fixed in 4% PFA for 12–24 h. For a subset of animals 5 out of 12 EXP mice , after PFA post-fixing, brains were cleared using a previously described iDISCO+ protocol<sup 64 https://www.nature.com/articles/s41586-026-11217-9 ref-CR64 </sup . Cleared brains were then imaged with light sheet microscopy at ×1.1 or ×1.3 magnification, and these images were registered to a common reference atlas as previously described<sup 65 https://www.nature.com/articles/s41586-026-11217-9 ref-CR65 </sup . We chose to register to the Paxinos Atlas<sup 66 https://www.nature.com/articles/s41586-026-11217-9 ref-CR66 </sup to map fibre tip placement to anatomically defined sub-areas of each SBN location, such as the lateral division of the periaqueductal grey or the NAc core and shell structures. Fibre tip locations were manually identified in registered light-sheet images and plotted together in the common Paxinos coordinate frame. Experimenters were blinded to group identity during histological verification. For another subset of animals 7 out of 12 EXP, 7 out of 9 OBS and 6 out of 8 NON mice , after PFA post-fixing, the extracted brains were cryoprotected in 30% sucrose and embedded in optimal cutting temperature mounting medium Fisher Healthcare Tissue-Plus OCT compound, Thermo Fisher Scientific for freezing over dry ice. Cryosections of the frozen tissue 30 μm slices were made and stamped directly onto glass microscope slides. The slices were washed with PBS, dried and then coverslipped with a mounting medium EMS Immuno Mount DAPI and DABSCO, Electron Microscopy Sciences, 17989-98, 180418 , and sealed with clear nail polish around the slide edges. After at least 12 h of drying, the slides were imaged with a digital robotic slide scanner NanoZoomer S60, C13210-01, Hamamatsu . Fibre placements for all recorded animals are summarized in Extended Data Fig. 4 https://www.nature.com/articles/s41586-026-11217-9 Fig9 , with coronal reference sections adapted from the Allen Reference Atlas<sup 66 https://www.nature.com/articles/s41586-026-11217-9 ref-CR66 </sup . DREADD-G<sub q</sub perturbation To assess the causal contribution of inhibitory populations to observation-induced changes in aggressive behaviour, a separate cohort of Vgat <sup +/− </sup OBS-G<sub q</sub or wild-type littermates Vgat <sup −/− </sup OBS-control animals expressing hM3Dq-mCherry in MeA, vLS, PMv and AH underwent an observation training regime consisting of five consecutive days of three resident–intruder observation sessions per day, using the same observer compartment setup described above. On the sixth day, the mice were subjected to a hard fight session against a CD1 aggressor screened a priori for aggression against other CD1 males. At 30 min before the hard fight, each animal received an intraperitoneal injection of clozapine- N -oxide CNO; 1 mg per kg; Hello Bio and was returned to its homecage to allow for sufficient drug bioavailability before the CD1 aggressor was introduced. Behavioural outcomes, using unsupervised clustering and annotation classification of fight outcome during the hard fight were compared between OBS-G<sub q</sub and OBS-control cohorts. Note that assignment to OBS-G<sub q</sub and OBS-control was determined by genotype; both groups received identical viral injections and CNO. Experimenters were not blinded to genotype during CNO administration of the hard fight. Statistical analysis All statistical tests were run in Python v.3.9.23 using numpy v.1.23.0 , pandas v.2.3.1 , matplotlib v.3.9.4 , seaborn v.0.13.2 , scipy v.1.13.1 , scikit-learn v.1.6.1 and statsmodels v.0.14.5 . Unless otherwise noted in the corresponding figure legend or Methods https://www.nature.com/articles/s41586-026-11217-9 Sec10 section, statistical analyses were conducted using biological replicates individual animals as the unit of replication. The exception is analysis of trial-level neural-behavioural covariation for example, decoding of trial-by-trial resident attack speed, subordinate speed; Fig. 3b–d https://www.nature.com/articles/s41586-026-11217-9 Fig3 , for which the trial is the appropriate unit of analysis given the trial-level nature of the scientific question being tested. Before any group comparison, samples were classified as independent or related based on experimental design: distinct animals assigned to different groups were treated as independent, and repeated measurements from the same animals across timepoints or conditions were treated as related. Sidedness of tests one-sided versus two-sided was determined a priori based on directional hypotheses specified before the corresponding analysis was performed, and is stated explicitly in the Methods https://www.nature.com/articles/s41586-026-11217-9 Sec10 section and figure legend for each analysis in which a one-sided test was used. In the absence of a specified a priori directional hypothesis, tests were two-sided by default. For comparisons between two independent groups, normality was assessed per group using the Shapiro–Wilk test scipy.stats.shapiro . For normally distributed data, equality of variance was assessed using Levene’s test scipy.stats.levene ; group means were compared using Student’s t -test scipy.stats.ttest ind when variances were equal, or Welch’s t -test scipy.stats.ttest ind, equal var=False when variances were unequal. For non-normally distributed data, the Brunner–Munzel test scipy.stats.brunnermunzel was used in place of the Mann–Whitney U -test as the primary nonparametric two-sample test. The Brunner–Munzel test was preferred because, in contrast to the Mann–Whitney U -test, it does not assume equal variances or distributional shape between the two groups being compared and remains valid under heteroscedasticity and unequal sample sizes—both common features of this dataset, where group sizes frequently differ due to histological exclusion criteria. In the rare cases in which the Brunner–Munzel test statistic was undefined for example, due to zero within-group variance , the Mann–Whitney U -test scipy.stats.mannwhitneyu was used as a fallback. Multiple comparisons were corrected using the Bonferroni or FDR Benjamini-Hochberg; statsmodels.stats.multitest.multipletests method as specified in the corresponding Methods https://www.nature.com/articles/s41586-026-11217-9 Sec10 section and figure legend. For comparisons between two related samples, the normality of the paired differences was assessed using the Shapiro–Wilk test. A paired t -test scipy.stats.ttest rel was used for normally distributed differences, and a Wilcoxon signed-rank test scipy.stats.wilcoxon was used otherwise. For comparisons across three independent groups EXP, OBS and NON , normality was assessed per group via the Shapiro–Wilk test and equality of variance via Levene’s test. For normally distributed data with equal variance, a one-way ANOVA scipy.stats.f oneway, equal var=True was used, followed by post hoc Student’s t -tests. For normally distributed data with unequal variance, Welch’s ANOVA scipy.stats.f oneway, equal var=False was used, followed by post hoc Welch’s t -tests. For non-normally distributed data, a Kruskal–Wallis test scipy.stats.kruskal was used, followed by post hoc Brunner–Munzel tests with Mann–Whitney U -test as a fallback, as described above . Post hoc pairwise comparisons were corrected using the Bonferroni or FDR method as specified in the corresponding figure legend. For the single analysis involving three related conditions day 1, day 8 and the hard fight, within the EXP group; Extended Data Fig. 1d https://www.nature.com/articles/s41586-026-11217-9 Fig6 , a Friedman test scipy.stats.friedmanchisquare was used in place of a repeated-measures ANOVA, followed by post hoc Wilcoxon signed-rank tests, corrected using the FDR method. For determining the difference between two flattened distributions NAc DA/LHb bias , a two-sample Kolmogorov–Smirnov test scipy.stats.ks 2samp was conducted. For assessing whether cross-validated decoding accuracy differed significantly from chance 0.5 , one-sample t -tests scipy.stats.ttest 1samp or Wilcoxon signed-rank tests scipy.stats.wilcoxon were applied to the distribution of per-fold accuracy scores pending normality, tested one-sided for above-chance decoding, with FDR correction across regions where applicable. Where LME models were fit for example, to test for a trend in unique variance across the training period; see ‘Decomposition of shared and unique variance in VMHvl activity across behavioural states’ section , convergence was assessed using up to six optimizer/estimation-method combinations in sequence L-BFGS, Powell and conjugate gradient, each under both restricted and full maximum likelihood , with the first converged fit retained; models for which no combination converged were excluded from the corresponding analysis. Convergence outcomes and the specific method used were logged for every fit as a diagnostic check on model reliability. Where session-level or pooled data were evaluated by permutation testing for example, pairwise transition-probability significance across UMAP clusters, decoding accuracy against chance and against group identity; Extended Data Fig. 1g–i https://www.nature.com/articles/s41586-026-11217-9 Fig6 , the number of permutations and null-construction procedures are described in the corresponding Methods https://www.nature.com/articles/s41586-026-11217-9 Sec10 section. Where multiple comparisons were corrected within a defined family rather than across all comparisons in a panel, family membership was established a priori and is stated explicitly in the corresponding figure legend. Family structure varied by analysis, reflecting the specific directional hypotheses under test. For behaviour cluster-level comparisons, clusters were divided into two a priori families—social clusters 1–5 and 7–9 and asocial clusters 6 and 10–13 —anticipating opposite directions of effect between the two. For Fig. 2 https://www.nature.com/articles/s41586-026-11217-9 Fig2 , families were defined within each experimental group EXP, OBS individually. For Fig. 3 https://www.nature.com/articles/s41586-026-11217-9 Fig3 , families were defined by valence NAc-DA, LHb and excitatory/inhibitory population identity. For Fig. 4 https://www.nature.com/articles/s41586-026-11217-9 Fig4 , families were defined by behavioural context training period, hard fight . For Fig. 5 https://www.nature.com/articles/s41586-026-11217-9 Fig5 , families were defined by cell type excitatory, inhibitory for all ARD-based analyses. Where a figure legend does not specify a family structure, correction was applied either within a single planned comparison no correction required or across the full set of comparisons presented in that panel. Our Article follows recommendations for inclusion and ethics in global research. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary https://www.nature.com/articles/s41586-026-11217-9 MOESM2 linked to this article. Data availability All processed data required to reproduce the analyses in this Article—including processed photometry traces, behaviour labels supervised and unsupervised and SLEAP-derived feature data—have been deposited at FigShare and are publicly accessible<sup 67 https://www.nature.com/articles/s41586-026-11217-9 ref-CR67 </sup https://doi.org/10.6084/m9.figshare.33860836 https://doi.org/10.6084/m9.figshare.33860836 . Raw acquisition data raw photometry recordings and raw video are available from the corresponding author on reasonable request. Source data https://www.nature.com/articles/s41586-026-11217-9 Sec26 are provided with this paper. Code availability Relevant code is available at GitHub https://github.com/FalknerLab/AggressionObservation https://github.com/FalknerLab/AggressionObservation . References 1. Olsson, A., Knapska, E. & Lindström, B. The neural and computational systems of social learning. Nat. Rev. Neurosci. 21 , 197–212 2020 . 2. Boone, T., Reilly, A. J. & Sashkin, M. Social learning theory. Group Organ. Stud. 2 , 384–385 1977 . 3. Bandura, A., Ross, D. & Ross, S. A. Vicarious reinforcement and imitative learning. J. Abnorm. Psychol. 67 , 601–607 1963 . 4. Yang, T. et al. Hypothalamic neurons that mirror aggression. Cell 186 , 1195–1211 2023 . 5. Vinograd, A., Nair, A., Kim, J. H., Linderman, S. W. & Anderson, D. J. Causal evidence of a line attractor encoding an affective state. Nature 634 , 910–918 2024 . 6. Li, X. et al. Violence exposure across multiple contexts as predictors of reactive and proactive aggression in Chinese preadolescents. Aggress. Behav. 48 , 319–330 2022 . 7. Wall Myers, T. D. et al. Understanding the link between exposure to violence and aggression in justice-involved adolescents. Dev. Psychopathol. 30 , 593–603 2018 . 8. Suzuki, H. & Lucas, L. R. Chronic passive exposure to aggression escalates aggressiveness of rat observers. Aggress. Behav. 36 , 54–66 2010 . 9. Williamson, C. M., Romeo, R. D. & Curley, J. P. Dynamic changes in social dominance and mPOA GnRH expression in male mice following social opportunity. Horm. Behav. 87 , 80–88 2017 . 10. Curley, J. P. Temporal pairwise-correlation analysis provides empirical support for attention hierarchies in mice. Biol. Lett. 12 , 20160192 2016 . 11. Willmore, L., Cameron, C., Yang, J., Witten, I. B. & Falkner, A. L. Behavioural and dopaminergic signatures of resilience. Nature 611 , 124–132 2022 . 12. Zhukovskaya, A. et al. Heightened lateral habenula activity during stress produces brainwide and behavioral substrates of susceptibility. Neuron https://doi.org/10.1016/j.neuron.2024.09.009 https://doi.org/10.1016/j.neuron.2024.09.009 2024 . 13. Rizzolatti, G. & Craighero, L. The mirror-neuron system. Annu. Rev. Neurosci. 27 , 169–192 2004 . 14. Lin, D. et al. Functional identification of an aggression locus in the mouse hypothalamus. Nature 470 , 221–226 2011 . 15. O’Connell, L. A. & Hofmann, H. A. The vertebrate mesolimbic reward system and social behavior network: a comparative synthesis. J. Comp. Neurol. 519 , 3599–3639 2011 . 16. Lo, L. et al. Connectional architecture of a mouse hypothalamic circuit node controlling social behavior. Proc. Natl Acad. Sci. USA 116 , 7503–7512 2019 . 17. Karigo, T. et al. Distinct hypothalamic control of same- and opposite-sex mounting behaviour in mice. Nature 589 , 258–263 2021 . 18. Golden, S. A. et al. Basal forebrain projections to the lateral habenula modulate aggression reward. Nature 534 , 688–692 2016 . 19. Kashtelyan, V., Lichtenberg, N. T., Chen, M. L., Cheer, J. F. & Roesch, M. R. Observation of reward delivery to a conspecific modulates dopamine release in ventral striatum. Curr. Biol. 24 , 2564–2568 2014 . 20. Stagkourakis, S., Spigolon, G., Liu, G. & Anderson, D. J. Experience-dependent plasticity in an innate social behavior is mediated by hypothalamic LTP. Proc. Natl Acad. Sci. USA 117 , 25789–25799 2020 . 21. Nordman, J. et al. Potentiation of divergent medial amygdala pathways drives experience-dependent aggression escalation. J. Neurosci. https://doi.org/10.1523/JNEUROSCI.0370-20.2020 https://doi.org/10.1523/JNEUROSCI.0370-20.2020 2020 . 22. de Moura Oliveira, V. E., Evrard, F., Faure, M. C. & Bakker, J. Social isolation and aggression training lead to escalated aggression and hypothalamus-pituitary-gonad axis hyperfunction in mice. Neuropsychopharmacology 49 , 1266–1275 2024 . 23. Yan, R. et al. The multi-stage plasticity in the aggression circuit underlying the winner effect. Cell 187 , 6785–6803.e18 2024 . 24. Falkner, A. L., Grosenick, L., Davidson, T. J., Deisseroth, K. & Lin, D. Hypothalamic control of male aggression-seeking behavior. Nat. Neurosci. 19 , 596–604 2016 . 25. Minakuchi, T. et al. Independent inhibitory control mechanisms for aggressive motivation and action. Nat. Neurosci. 27 , 702–715 2024 . 26. Wong, L. C. et al. Effective modulation of male aggression through lateral septum to medial hypothalamus projection. Curr. Biol. 26 , 593–604 2016 . 27. Barbier, M. et al. Projections from the dorsomedial division of the bed nucleus of the Stria terminalis to hypothalamic nuclei in the mouse. J. Comp. Neurol. https://doi.org/10.1002/cne.24988 https://doi.org/10.1002/cne.24988 2020 . 28. Yamamoto, R., Ahmed, N., Ito, T., Gungor, N. Z. & Pare, D. Optogenetic study of anterior BNST and basomedial amygdala projections to the ventromedial hypothalamus. eNeuro 5 , ENEURO.0204-18.2018 2018 . 29. Hong, W., Kim, D.-W. & Anderson, D. J. Antagonistic control of social versus repetitive self-grooming behaviors by separable amygdala neuronal subsets. Cell 158 , 1348–1361 2014 . 30. Pereira, T. D. et al. SLEAP: A deep learning system for multi-animal pose tracking. Nat. Methods 19 , 486–495 2022 . 31. Koolhaas, J. M. et al. Coping styles in animals: current status in behavior and stress-physiology. Neurosci. Biobehav. Rev. 23 , 925–935 1999 . 32. Klibaite, U., Berman, G. J., Cande, J., Stern, D. L. & Shaevitz, J. W. An unsupervised method for quantifying the behavior of paired animals. Phys. Biol. 14 , 015006 2017 . 33. Guthman, E. M., Iravedra-Garcia, J. M., Sirrs, L., Le, A. & Falkner, A. L. Context rescales a social action code in a hormone-sensitive network. Preprint at bioRxiv https://doi.org/10.1101/2025.09.24.678271 https://doi.org/10.1101/2025.09.24.678271 2025 . 34. Sych, Y., Chernysheva, M., Sumanovski, L. T. & Helmchen, F. High-density multi-fiber photometry for studying large-scale brain circuit dynamics. Nat. Methods 16 , 553–560 2019 . 35. Kim, C. K. et al. Simultaneous fast measurement of circuit dynamics at multiple sites across the mammalian brain. Nat. Methods 13 , 325–328 2016 . 36. Yang, B., Karigo, T. & Anderson, D. J. Transformations of neural representations in a social behaviour network. Nature 608 , 741–749 2022 . 37. Sun, F. et al. A genetically encoded fluorescent sensor enables rapid and specific detection of dopamine in flies, fish, and mice. Cell 174 , 481–496 2018 . 38. Williams, A. H. et al. Discovering precise temporal patterns in large-scale neural recordings through robust and interpretable time warping. Neuron 105 , 246–259 2020 . 39. Yamaguchi, T. & Lin, D. Functions of medial hypothalamic and mesolimbic dopamine circuitries in aggression. Curr. Opin. Behav. Sci. 24 , 104–112 2018 . 40. Adjei, M. P. et al. Familiarity gates socially transmitted aggression via the medial amygdala. J. Neurosci. 45 , e1018252025 2025 . 41. Dupré la Tour, T., Eickenberg, M., Nunez-Elizalde, A. O. & Gallant, J. L. Feature-space selection with banded ridge regression. Neuroimage 264 , 119728 2022 . 42. Roth, B. L. DREADDs for neuroscientists. Neuron 89 , 683–694 2016 . 43. Berkowitz, L. Aggression: Its Causes, Consequences, and Control McGraw-Hill, 1993 . 44. Lagerspetz, K. M. J. Aggression and aggressiveness in laboratory mice. In Aggressive Behaviour eds Garattini, S. & Sigg, E. B. 77–85 Excerpta Medica Foundation, 1969 . 45. Hamid, A. A. et al. Mesolimbic dopamine signals the value of work. Nat. Neurosci. 19 , 117–126 2016 . 46. Warren, B. L. et al. Neurobiological sequelae of witnessing stressful events in adult mice. Biol. Psychiatry 73 , 7–14 2013 . 47. Nakatake, Y. et al. The effects of emotional stress are not identical to those of physical stress in mouse model of social defeat stress. Neurosci. Res. 158 , 56–63 2020 . 48. Sial, O. K., Warren, B. L., Alcantara, L. F., Parise, E. M. & Bolaños-Guzmán, C. A. Vicarious social defeat stress: bridging the gap between physical and emotional stress. J. Neurosci. Methods 258 , 94–103 2016 . 49. Tomasello, M. Emulation learning and cultural learning. Behav. Brain Sci. 21 , 703–704 1998 . 50. Mondoloni, S. et al. Serotonin release in the habenula during emotional contagion promotes resilience. Science 385 , 1081–1086 2024 . 51. Allen Reference Atlas—Mouse Brain https://atlas.brain-map.org https://atlas.brain-map.org Allen Institute for Brain Science, 2011 . 52. Vong, L. et al. Leptin action on GABAergic neurons prevents obesity and reduces inhibitory tone to POMC neurons. Neuron 71 , 142–154 2011 . 53. Lee, H. et al. Scalable control of mounting and attack by Esr1 <sup +</sup neurons in the ventromedial hypothalamus. Nature 509 , 627–632 2014 . 54. Friard, O. & Gamba, M. BORIS: a free, versatile open-source event-logging software for video/audio coding and live observations. Methods Ecol. Evol. 7 , 1325–1330 2016 . 55. Bohnslav, J. P. et al. DeepEthogram, a machine learning pipeline for supervised behavior classification from raw pixels. eLife 10 , e63377 2021 . 56. Sun, J. J. et al. The multi-agent behavior dataset: mouse dyadic social interactions. Adv. Neural Inf. Process. Syst. 2021 , 1–15 2021 . 57. Woo, S., Park, J., Lee, J.-Y. & Kweon, I. S. CBAM: convolutional block attention module. In Computer Vision—ECCV 2018 3–19 Springer, 2018 . 58. Fang, Y.-Y., Yamaguchi, T., Song, S. C., Tritsch, N. X. & Lin, D. A hypothalamic midbrain pathway essential for driving maternal behaviors. Neuron 98 , 192–207 2018 . 59. Guthman, E. M. et al. Two-color multifiber photometry recordings of the social behavior network in mice. J. Vis. Exp. 229 , e70049 2026 . 60. Zhang, Z.-M., Chen, S. & Liang, Y.-Z. Baseline correction using adaptive iteratively reweighted penalized least squares. Analyst 135 , 1138–1146 2010 . 61. Lundberg, S. & Lee, S.-I. A unified approach to interpreting model predictions. Preprint at https://doi.org/10.48550/arXiv.1705.07874 https://doi.org/10.48550/arXiv.1705.07874 2017 . 62. Kobak, D. et al. Demixed principal component analysis of neural population data. eLife 5 , e10989 2016 . 63. Semedo, J. D., Zandvakili, A., Machens, C. K., Yu, B. M. & Kohn, A. Cortical areas interact through a communication subspace. Neuron 102 , 249–259 2019 . 64. Pisano, T. J. et al. Homologous organization of cerebellar pathways to sensory, motor, and associative forebrain. Cell Rep. 36 , 109721 2021 . 65. Renier, N. et al. Mapping of brain activity by automated volume analysis of immediate early genes. Cell 165 , 1789–1802 2016 . 66. Chon, U., Vanselow, D. J., Cheng, K. C. & Kim, Y. Enhanced and unified anatomical labeling for a common mouse brain atlas. Nat. Commun. 10 , 5067 2019 . 67. Iravedra-García, J. M. et al. Data for ‘Aggression experience and observation drive shared behavioural and neural changes’. Figshare https://doi.org/10.6084/m9.figshare.33860836 https://doi.org/10.6084/m9.figshare.33860836 2026 . Acknowledgements We thank J. Shaevitz, M. Murthy, I. Witten, C. Pena and the members of the Falkner laboratories for useful discussions; S. Oline for assistance with camera setup and syncing; S. Yokoyama for help with fibre implants; A. Le for help genotyping; the staff at the PNI Viral Core Facility for reagents and the PNI Brain Registration and Histology Core Facility for assistance with histology; and T. Yamaguchi for providing us with the cre-excluding viral constructs. Funding Funding was from DP2MH126375 A.L.F. , NIH R01MH126035 A.L.F. , New York Stem Cell Foundation A.L.F. , Simons Collaboration for the Global Brain A.L.F. , Klingenstein Foundation A.L.F. , the Alfred P. Sloan Fellowship A.L.F. and NIH K99MH135212 E.M.G. . A.L.F. is a New York Stem Cell Foundation Robertson Investigator. Ethics declarations Competing interests The authors declare no competing interests. Peer review Peer review information Nature thanks Adi Nair, Moriel Zelikowsky and the other, anonymous, reviewer s for their contribution to the peer review of this work. Additional information Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Extended data figures and tables Extended Data Fig. 1 Additional behavioural and classification metrics across the training period and hard fight. https://www.nature.com/articles/s41586-026-11217-9/figures/6 a , Left: Confusion matrix, multilayer perceptron trained on SLEAP-derived features predicting manually annotated attack behaviour balanced accuracy = 0.92 . Right: Binary prediction of attack behaviours red against hand-annotated ground truth black , representative example sessions. b , Occupancy left , latency to first attack middle , and persistence average consecutive attack-bout length, right , early first 2 days versus late last 2 days sessions of the training period for EXP animals n = 10 paired animals . Box plots: line represents median; box, 25th–75th percentile; whiskers, min/max within 1.5x IQR; points overlaid; lines connect the same animal across early and late sessions. Two-sided paired t-tests or Wilcoxon signed-rank tests: occupancy paired t-test, t 9 = 0.702, p = 0.5003, n.s. ; latency Wilcoxon signed-rank, W = 6.000, p = 0.0273 ; persistence paired t-test, t 9 = 0.976, p = 0.3547, n.s. . c , Occupancy of attacks across days of the training period for EXP orange, N = 10 and OBS magenta, N = 9 animals mean ± SEM . Significance brackets indicate days where EXP and OBS occupancy differ significantly two-sided t-tests or Brunner-Munzel tests, FDR-corrected across all days; D1: Brunner-Munzel, BM = 2.808, p = 0.0305 ; D2: Brunner-Munzel, BM = 4.289, p = 0.0059 ; D3: t-test, t 17 = −2.937, p = 0.0246 ; D4: Brunner-Munzel, BM = 4.216, p = 0.0059 ; D5: t-test, t 17 = −2.666, p = 0.0305 ; D6: Brunner-Munzel, BM = 2.308, p = 0.0393 ; D7: t-test, t 17 = −2.499, p = 0.0307 ; D8: t-test, t 17 = −1.745, p = 0.0991, n.s. . d , Occupancy of attacks during day 1, day 8, and the hard fight for EXP animals n = 10 paired animals; box plots as in b, with lines connecting the same animal across the three timepoints . Omnibus comparison via Friedman test F = 1.400, p = 0.4966, n.s. e , Occupancy left and latency right of fleeing events during the hard fight by EXP orange, N = 10 , OBS magenta, N = 9 , and NON grey, N = 8 animals. Occupancy: omnibus Kruskal-Wallis, H = 7.652, p = 0.0218; post-hoc two-sided Brunner-Munzel tests, FDR-corrected across comparisons: EXP vs. OBS BM = 1.207, p = 0.2440, n.s. , EXP vs. NON BM = 3.461, p = 0.0100 , OBS vs. NON BM = 2.150, p = 0.0742, n.s. . Latency: omnibus Kruskal-Wallis, H = 1.662, p = 0.4356, n.s. Box plots as in b . f , Probability distribution of frames containing attacks found in each UMAP cluster during the hard fight. Clusters 2, 3, and 4 contain the highest proportion of attack frames, consistent with their designation as conflict clusters. g , Leave-one-session-out cross-validation balanced accuracy of group ID EXP, OBS, NON predictions using UMAP cluster occupancy as the classification feature, evaluated against a null distribution generated by a permutation test 1,000 permutations, group labels shuffled independently per session, one-sided; p = 0.0010 . h , Same as g but using pairwise transition probabilities between UMAP clusters as the classification feature permutation test as in g; p = 0.0020 . i , Top : Same as g but using cluster persistence average consecutive bout duration, seconds as the classification feature permutation test as in g ; p = 0.0010 . Bottom: Persistence of all 13 behaviour clusters across groups EXP N = 10, OBS N = 9, NON N = 8 during the hard fight box plots as in b ; one-way ANOVA or Kruskal-Wallis test pending normality, post-hoc two-sided t-tests or Brunner-Munzel tests, FDR-corrected; full statistics in Table S1 https://www.nature.com/articles/s41586-026-11217-9 MOESM3 . j , Subset of averaged SLEAP-derived features across groups during the hard fight EXP N = 10, OBS N = 9, NON N = 8; box plots as in b . Features shown: resident-intruder centroid distance, resident centroid speed, resident-intruder body overlap, intruder centroid speed. Full statistics in Table S1 https://www.nature.com/articles/s41586-026-11217-9 MOESM3 . For all comparisons, p < 0.05, p < 0.01, p < 0.001. Extended Data Fig. 2 Enumeration and description of SLEAP behaviour features and unsupervised behaviours. https://www.nature.com/articles/s41586-026-11217-9/figures/7 a , Schematic and mathematical descriptions of all SLEAP-derived features used for behavioural segmentation, organized by category. Distance features capture inter-animal proximity and approach dynamics. The bounding box category includes intersection over union IoU , computed via Shapely polygon operations, which measures the degree of physical overlap between animal bounding boxes. Social orientation features capture the directed facing angle of each animal relative to the other. Postural features include within-mouse head angle centroid–neck–head angle, indexing sniffing posture and body curvature and within-mouse body length tailbase-to-head distance, indexing body elongation . Speed features are computed as the rate of change ROC of position over a 4-frame window 100 ms , including within-mouse translational centroid speed and head-specific speed for both resident and intruder, as well as a summed speed feature aggregating across temporal lags spanning ±2 s. b , Unsupervised behaviour states resulting from watershed segmentation of the UMAP embedding, with descriptions organized according to the unsupervised hierarchical clustering results depicted in Fig. 1m https://www.nature.com/articles/s41586-026-11217-9 Fig1 , which segregates behavioural states into three discrete categories: conflict and aggression-preparatory states red; clusters 1–4 , investigative or social engagement states gold; clusters 5, 7–9 , and disengaged or avoidant states green; clusters 6, 10–13 . The behavioural content of each cluster is ascertained via a combination of manual curation of video frames across multiple sessions and animals and feature coding as shown in c . c , Feature maps of all 13 behavioural clusters across all recorded SLEAP-derived behaviour features, including resident and intruder centroid speed, resident and intruder head speed, resident-intruder centroid distance, resident head to intruder head distance, resident head to intruder tailbase distance, intruder head to resident tailbase distance, resident-intruder orientation, intruder-resident orientation, resident-intruder body overlap IoU , resident body length, intruder body length, resident body orientation, and intruder body orientation. Each map depicts the mean z-scored feature value projected onto the UMAP behaviour space, averaged across animals and sessions. d , Representative video frames depicting examples of each behavioural cluster, organized according to the same hierarchical grouping as in b . Conflict states clusters 1–4 : mutual conflict, resident aggression, aggression preparation 1, aggression preparation 2. Investigative states clusters 5, 7–9 : huddled investigation, non-specific/orofacial investigation, mutual investigation, mutually facing investigation. Disengaged/avoidant states clusters 6, 10, 11, 13 : intruder disengages, intruder approaches/investigates, avoidance with resident movement 1, avoidance with resident movement 2, avoidance with resident scrunch/groom. Resident animals are indicated by cyan markers and intruder animals by red markers. Extended Data Fig. 3 Nonsocial exposure is insufficient to recapitulate the behavioural changes induced by aggression observation. https://www.nature.com/articles/s41586-026-11217-9/figures/8 a , Schematic depicting the timeline of exposure training. In contrast to OBS animals, nonsocial exposure XPO, N = 5 animals observed non-aggressive male cagemates for 8 days, and on day 9 were pitted against three rounds of CD1 aggressors. b , Leave-one-session-out cross-validation balanced accuracy of group ID OBS vs. XPO predictions using occupancy of UMAP clusters left , persistence mean length of cluster bouts, middle , and pairwise cluster transition probabilities right as classification features. For each metric, test balanced accuracy was evaluated against a null distribution generated by a permutation test 1,000 permutations, group labels shuffled independently per session, one-sided; occupancy, p = 0.0010 ; persistence, p = 0.0010 ; transitions, p = 0.0010 . c , Occupancy of a subset of behaviour clusters for OBS N = 9 and XPO N = 5 animals during the hard fight: clusters 2 and 4 conflict , 5 and 7 investigative , and 6 and 10 asocial — two representative clusters per behavioural category, selected a priori based on the category structure established in Fig. 1 https://www.nature.com/articles/s41586-026-11217-9 Fig1 . Box plots: line represents median; box, 25th–75th percentile; whiskers, min/max within 1.5x IQR; points overlaid; individual points overlaid. Group differences evaluated via two-sided t-tests or Brunner-Munzel tests where no directional hypothesis existed, and one-sided tests where an a priori direction was established by the corresponding effect in Fig. 1 https://www.nature.com/articles/s41586-026-11217-9 Fig1 : clusters 2, 4, 5, 7 tested one-sided OBS XPO followed by FDR correction; clusters 6, 10 tested one-sided OBS < XPO followed by FDR correction. Full statistics in Table S1 https://www.nature.com/articles/s41586-026-11217-9 MOESM3 . d , Aggregate transition weights between behaviour-category clusters as in Fig. 1o https://www.nature.com/articles/s41586-026-11217-9 Fig1 , OBS versus XPO, hard fight, for a subset of transitions of a priori interest from Fig. 1 https://www.nature.com/articles/s41586-026-11217-9 Fig1 : conflict→conflict, conflict→investigation and asocial→investigation one-sided, OBS XPO ; investigation→investigation and investigation→asocial two-sided ; and asocial→asocial one-sided, OBS < XPO . N and box plots as in c . Tests: one-sided t-tests or Brunner-Munzel tests, followed by FDR corrections across comparisons in the same family of origin behaviours. Full statistics in Table S1 https://www.nature.com/articles/s41586-026-11217-9 MOESM3 . e , Classification strategy for determining EXP-like or NON-like behavioural tendency of OBS versus XPO animals during the hard fight. A support vector machine SVM classifier is trained on EXP and NON behaviour using cluster occupancy, persistence, or pairwise transition probabilities as separate feature sets. Each classifier is then applied to OBS and XPO animals, with individuals classified as EXP-like or NON-like based on the classifier output. f , Left : Distribution of P classified as EXP for OBS N = 9, magenta versus XPO N = 5, cyan animals, using cluster occupancy as the classifier feature two-sided Brunner-Munzel test, BM = −2.956, p = 0.0193 . Box plots as in c. Middle : as in left, using cluster persistence BM = −3.164, p = 0.0083 . Right : as in left, using pairwise transition probabilities BM = −2.563, p = 0.0281 . For all comparisons, p < 0.05, p < 0.01, p < 0.001. Extended Data Fig. 4 Summary of optical fibre locations and viral expression across all recorded animals. https://www.nature.com/articles/s41586-026-11217-9/figures/9 a , Representative GCaMP6f green or JRCaMP2b red expression across all 12 recorded brain areas. Scale bars 500 μm. Representative of fibre placement confirmed via histology in the majority of subjects per region range: 8/12–12/12 for EXP; 9/9 for OBS; 8/8 for NON, across all regions ; full per-region counts in b. b , Accepted fibre placements for every experimental subject across all 12 recorded brain areas. Coronal schematics are adapted from the Allen Brain Atlas<sup 51 https://www.nature.com/articles/s41586-026-11217-9 ref-CR51 </sup . Fraction of subjects per group with histologically confirmed placement, by region EXP / OBS / NON : PrL, 11/12 / 9/9 / 8/8; vLS, 12/12 / 9/9 / 8/8; POA, 11/12 / 9/9 / 8/8; BNST, 10/12 / 9/9 / 8/8; AH, 10/12 / 9/9 / 8/8; MeA, 11/12 / 9/9 / 8/8; VMH, 12/12 / 9/9 / 8/8; PAG, 8/12 / 9/9 / 8/8; PMv, 9/12 / 9/9 / 8/8; LHb excitatory only , 9/12 / 9/9 / 8/8; PA, 9/12 / 9/9 / 8/8; NAc dopamine , 11/12 / 9/9 / 8/8. Extended Data Fig. 5 Supervised classification of observer behaviour: model architecture, behaviour dynamics, and neural decoding. https://www.nature.com/articles/s41586-026-11217-9/figures/10 a , Model architecture. Input clips of 20 frames 0.5 s are passed through stochastic augmentations during training, followed by four 3D convolutional attention blocks, each consisting of a 3D convolution, batch normalization, max pooling, and a spatial attention module. The encoder output is pooled into a 512-dimensional feature vector and passed to a three-layer MLP classifier producing softmax probabilities over seven behavioural classes. The seven output classes are merged into four behaviour categories: attention attention , self-directed grooming, scratching , idle still , and exploratory investigate, rearing, turning . Fine-grained class colours are shown within each category in the legend. b , Spatial attention maps from the first three encoder blocks for four behavioural classes selected to span distinct postures rearing, scratching, turning, attention . Each row shows a single representative frame from one held-out clip. Maps from the fourth block are omitted due to coarse spatial resolution. Within each panel, attention values are min-max normalized; brighter colours indicate higher attention weight. c , Row-normalized confusion matrix for the four behaviour categories, averaged across the 5 cross-validation folds. Cell values give the mean percentage of ground-truth samples assigned to each predicted category ± SD across folds. d , Example behavioural time series from three training sessions across the first 5 min of observation. Coloured bars indicate the classifier assigned behavioural category at each frame: attention blue , self-directed yellow , idle grey , and exploratory orange . Dark red dots mark demonstrator attack labels from an independent attack behaviour classifier see Supervised behaviour classification in Methods https://www.nature.com/articles/s41586-026-11217-9 Sec10 . e , Behaviour category occupancy of OBS animals N = 9 mice during training period. f , Daily behaviour category occupancy for OBS top, N = 9 mice and XPO bottom, N = 5 mice animals across the training period. Attention proportion was modelled separately per condition by a linear mixed-effects model with day as a continuous fixed effect and a random intercept per mouse occupancy ~ day + 1 | mouse , REML ; significance is the two-sided Wald z-test on the day coefficient OBS, β = 0.018 occupancy/ day, p = 0.002 ; XPO, β = 0.011, p = 0.229, n.s. . Error bars, ±SEM across mice. g , Behaviour category occupancy fraction of time occupied relative to total session length compared between OBS magenta, N = 9 and XPO cyan, N = 5 animals during the training period. From left to right: attention occupancy two-sided independent samples t-test: t 12 = 2.26, p = 0.044 , self-directed occupancy two-sided independent samples t-test: t 12 = 0.166, p = 0.8713, n.s. , idle occupancy two-sided Brunner-Munzel test: BM = 1.884, p = 0.0852 n.s. , exploratory occupancy two-sided independent samples t-test: t 12 = −2.47, p = 0.0294 . Box plots: line represents median; box, 25th–75th percentile; whiskers, min/max within 1.5x IQR; individual animal points overlaid h , Entropy bits of behaviour category occupancy distributions for OBS and XPO animals, computed as Shannon entropy across the four merged behavioural categories two-sided independent samples t-test: t 12 = 2.248, p = 0.062 n.s. . Box plot and N as in g. i , Behaviour occupancy of OBS animals N = 9 across attack relative time windows compared to a matched per-animal baseline. Windows: before attack onset −25 to 0% , attack onset 0 to 25% , attack middle 25 to 75% , attack end 75 to 100% , and after attack offset 100 to 125% ; baseline is mean behaviour occupancy across all training days. Error bars, ±SEM. Asterisks indicate significant deviation from baseline two-sided paired t-tests with FDR correction, for full statistics see Table S1 https://www.nature.com/articles/s41586-026-11217-9 MOESM3 . j , Left : Average difference in behaviour category occupancy between attack and control periods across OBS animals N = 9 . Control segments were randomly drawn from non-attack frames across 50 independent trials, matched in total duration to attack segments two-sided paired t-tests with FDR correction; attention: 0.0063 . Self-directed: 0.0083 . Right : Average mean latency to onset for each behavioural state during attack versus control periods across OBS animals N = 9 . Error bars, ±SEM across trials two-sided paired t-tests with FDR correction; n.s . k , Average behaviour category occupancy at attack onset, before vs. after time-shifting across OBS animals N = 9 . A window of 2 s around attack onset is used for this analysis. After time-shifting, there is a significant increase in attention p = 0.0098 and decrease in idle p = 0.0098 , and exploratory behaviours p = 0.0372 two-sided paired t-tests with FDR correction . Error bars, ±SEM. l , Decoding performance F1 score for behaviour category labels in OBS animals N = 9 , comparing all recorded regions combined filled bars , the best single region outlined bars , and permutation-based chance grey bars . The best single region was selected independently for each mouse and behaviour as the region yielding the highest mean F1 across cross-validation repeats. Error bars indicate ± SEM; dots show individual mice. Upper brackets indicate all regions vs. chance paired comparisons, lower brackets indicate all regions vs. best single region paired comparisons. P-values of two-sided paired t-tests with FDR correction are reported as all vs. chance, all vs. best single region : attention p = 9.54e-9 , p = 7.65e-4 ; self-directed p = 1.48e-6 , p = 1.87e-6 ; idle p = 3.28e-4 , p = 2.02e-4 ; exploratory p = 6.01e-7 , p = 8.89e-6 . m , Average SHAP SHapley Additive exPlanations values of signals’ contribution to decoding performance across OBS animals N = 9 , split by observer behaviour category. Error bars, ±SEM. For all comparisons, p < 0.05, p < 0.01, p < 0.001. Extended Data Fig. 6 Overview of the time-shifting pipeline and associated controls. https://www.nature.com/articles/s41586-026-11217-9/figures/11 a , Overview schematic of the time-shifting process. Top left : A tensor of SBN activity is constructed with dimensions trial × attack window × neural population. Top middle: The model learns a population-specific activity template from the components of each recorded population. Bottom middle: The model learns the likely timing of activity on a trial-by-trial basis, yielding a distribution of per-trial time shifts ranging from −2 s to +2 s. Trials shifted in the negative direction are depicted in dark green and trials shifted in the positive direction in light green; unshifted trials are shown in grey-green. A negative skew in shift distribution shifts <0 is evaluated via a one-sided Wilcoxon signed-rank test N = 1,115 trials, W = 258,136.5, p = 3.98 × 10<sup −7</sup . Top right : The end result of the pipeline proceeds from unshifted signals through time-shifted alignment to attention-state filtering ATTN+ versus ATTN− . b , Peri-event time histograms of excitatory top and inhibitory bottom SBN populations and valence areas for OBS animals, comparing time-shifted, attention-filtered magenta versus unshifted, attention-filtered blue attack-aligned signals z-ΔF/F, mean, shaded area ± SEM . Shifted and unshifted signals are shown for each recorded region across the full network. For each population, post-attack mean activity baseline-corrected is compared against zero using two-sided one-sample t-tests or Wilcoxon signed-rank tests pending normality, with FDR correction across all regions. For full statistics, including N per population, see Table S1 https://www.nature.com/articles/s41586-026-11217-9 MOESM3 . For all comparisons, p < 0.05, p < 0.01, p < 0.001. c , To validate that time-shifting of neural signals does not spuriously generate neural activity, three control conditions were devised in which the same time-shifting model was applied to OBS animals. In the first condition, the model was applied to neural signals during randomly resampled behavioural events generated by a Markov chain trained to reproduce attack patterns frequency and duration at random time points not overlapping with real aggression “synthetic,” light blue . In the second condition, the model was applied to neural signals during high resident velocity events unrelated to aggression “speed,” dark blue-green . In the third condition, the model was applied to NON animal neural signals during random indices derived from real attack timepoints in OBS animals “yoked,” grey . d , Peri-event time histograms z-ΔF/F of SBN and valence system activity for synthetic left and speed right control conditions mean, shaded area ± SEM . Neural signals from each control condition are overlaid on OBS signals from Fig. 2j https://www.nature.com/articles/s41586-026-11217-9 Fig2 . Extended Data Fig. 7 Attack-aligned neural signal decay dynamics across aggression experience and observation. https://www.nature.com/articles/s41586-026-11217-9/figures/12 a , Illustration of the half-decay time metric applied to representative EXP left and OBS right subjects for VMHvl excitatory and inhibitory populations. Each trace shows the baseline-corrected, attack-aligned mean z-ΔF/F signal. The peak value red dot is identified within the post-attack window and the half-decay time t½ yellow dot is defined as the time from peak to the point at which the signal first returns to 50% of its peak value, estimated with sub-frame precision via linear interpolation. The baseline noise floor 2 × baseline SD is used as a minimum threshold for valid peak detection. b , Half-decay time t½ s of attack-aligned calcium signals across all recorded excitatory green and inhibitory blue SBN populations, for EXP orange and OBS magenta animals during the training period. Box plots: line represents median; box, 25th–75th percentile; whiskers, min/max within 1.5x IQR; individual animal points overlaid. Regions are ordered with excitatory populations first, followed by inhibitory populations. For each region, EXP versus OBS comparisons are performed using two-sided independent samples t-tests or Brunner-Munzel tests pending normality, with FDR correction across all populations. For detailed statistics, see Table S1 https://www.nature.com/articles/s41586-026-11217-9 MOESM3 . c , Within-group comparison of half-decay time t½ between excitatory light green and inhibitory light blue populations for EXP animals, shown for each paired SBN region VMHvl, PMv, AH, POA, MeA, PA, BNST, lPAG, vLS, PrL . Each point represents one animal; lines connect paired excitatory and inhibitory measurements from the same animal. Significance brackets indicate regions where excitatory half-decay time is significantly shorter than inhibitory half-decay time one-sided paired t-tests or Wilcoxon signed-rank tests pending normality, with FDR correction across all regions . For detailed statistics, see Table S1 https://www.nature.com/articles/s41586-026-11217-9 MOESM3 . d , As in c but for OBS animals. For all comparisons, p < 0.05, p < 0.01, p < 0.001. Extended Data Fig. 8 Behaviour feature encoding during the hard fight. https://www.nature.com/articles/s41586-026-11217-9/figures/13 a , Average group-specific tuning curves mean, shaded area +/− SEM depicting the encoding of inter-animal proximity across hypothalamic and amygdala regions during the hard fight, shown as a function of binned proximity close to far . Each tuning curve represents a set of GLM weights fit for each feature and each region independently, using a leave-one-animal-out cross-validation procedure, for EXP orange , OBS magenta , and NON grey animals. Boxed regions indicate representative neural populations selected for display. Highlighted areas in each curve denote the first five bins blue and last five bins red of each tuning curve, corresponding to low and high values of the encoded feature respectively. Tuning curves for which the cross-validated R² does not exceed zero are blurred to indicate poor model fit and low predictive power; region–feature combinations are only included in group comparisons if at least two groups show R² significantly greater than zero one-sided Wilcoxon signed-rank test . For full R² comparisons and pairwise group statistics, see Table S1 https://www.nature.com/articles/s41586-026-11217-9 MOESM3 . b , As in a but for resident self-speed low to high . c , As in a but for CD1 aggressor speed low to high . Extended Data Fig. 9 Neural correlates of conflict during the hard fight. https://www.nature.com/articles/s41586-026-11217-9/figures/14 a , Peri-event time histograms z-ΔF/F of inhibitory PMv, AH, MeA, and BNST populations, aligned to conflict bout onset irrespective of attack direction during the hard fight, for EXP orange , OBS magenta , and NON grey animals mean ± SEM . b , Mean z-ΔF/F during the conflict period for each recorded population, excitatory populations and LHb top row and inhibitory populations and NAc DA bottom row , EXP/OBS/NON. Individual animals shown as points; horizontal line indicates the group mean, error bar indicates +/− SEM. N per population in Table S1 https://www.nature.com/articles/s41586-026-11217-9 MOESM3 . Group comparisons: omnibus one-way ANOVA or Kruskal-Wallis test pending normality, post-hoc two-sided t-tests or Brunner-Munzel tests, FDR-corrected within population; full statistics in Table S1 https://www.nature.com/articles/s41586-026-11217-9 MOESM3 . For all comparisons, p < 0.05, p < 0.01, p < 0.001. Extended Data Fig. 10 Inhibitory ensemble state predicts aggression preparedness and is shaped by social experience. https://www.nature.com/articles/s41586-026-11217-9/figures/15 Schematic summarizing the relationship between inhibitory SBN ensemble activity and behavioural state across experimental groups. The x-axis represents the inhibitory ensemble state, indexed by population activity level and/or VMHvl coupling strength. The y-axis represents aggression preparedness, ranging from avoidant to aggressively prepared. NON animals, which lack social experience, occupy a high inhibitory ensemble / low preparedness state. OBS and EXP animals shift rightward and upward along both axes following observation and direct fighting experience respectively, reflecting reduced inhibitory tone and increased aggression preparedness. Reactivation of inhibition returns OBS animals toward the NON-like state, demonstrating that inhibitory ensemble activity is both necessary and sufficient to gate aggression preparedness in a novel aggression context. Supplementary information Supplementary Information download PDF https://static-content.springer.com/esm/art%3A10.1038%2Fs41586-026-11217-9/MediaObjects/41586 2026 11217 MOESM1 ESM.pdf Legend for Supplementary Tables 1 and 2, and a description of the source data files. Supplementary Table 1 download XLSX https://static-content.springer.com/esm/art%3A10.1038%2Fs41586-026-11217-9/MediaObjects/41586 2026 11217 MOESM3 ESM.xlsx Statistical results underlying comparisons summarized in the Figures and Extended Data Figures, including test details not otherwise reported in the corresponding legend. Supplementary Video 1 download MP4 https://static-content.springer.com/esm/art%3A10.1038%2Fs41586-026-11217-9/MediaObjects/41586 2026 11217 MOESM4 ESM.mp4 Example behaviour during aggression experience, observation and non-social control training. Example behaviour for aggression experience EXP, left , observation OBS, middle and non-social control NON, right . An observer animal in OBS is positioned on a platform to the right and above the interacting animals, and is also filmed using an observer camera. Supplementary Video 2 download MP4 https://static-content.springer.com/esm/art%3A10.1038%2Fs41586-026-11217-9/MediaObjects/41586 2026 11217 MOESM5 ESM.mp4 Example behaviours representative of unsupervised behavioural clusters. Example behaviours from unsupervised clusters Extended Fig. 2 . Supplementary Video 3 download MP4 https://static-content.springer.com/esm/art%3A10.1038%2Fs41586-026-11217-9/MediaObjects/41586 2026 11217 MOESM6 ESM.mp4 Example synchronous multi-signal fibre photometry recording during aggression experience. Example synchronous recording of excitatory red , inhibitory blue and dopaminergic green signals in the SBN during aggression experience. Rights and permissions Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author s and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ http://creativecommons.org/licenses/by-nc-nd/4.0/ . About this article Cite this article Iravedra-García, J.M., Guthman, E.M., Tsin, D. et al. Aggression experience and observation drive shared behavioural and neural changes. Nature 2026 . https://doi.org/10.1038/s41586-026-11217-9 - Received: - Accepted: - Published: - Version of record: - DOI: https://doi.org/10.1038/s41586-026-11217-9