Paper deep dive

Social Reality Construction via Active Inference: Modeling the Dialectic of Conformity and Creativity

Kentaro Nomura, Takato Horii

Year: 2026Venue: arXiv preprintArea: cs.MAType: PreprintEmbeddings: 47

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%

Last extracted: 4/14/2026, 1:37:46 AM

Summary

The paper introduces a multi-agent simulation model based on active inference to study the dialectical construction of social reality. By integrating internal generative models with creative artifact production and selective communication via the Metropolis–Hastings Naming Game, the model demonstrates how informationally cohesive social groups emerge endogenously and how individual creative acts mutually constitute social representations and observation distributions.

Entities (4)

Active Inference · framework · 100%Metropolis–Hastings Naming Game · algorithm · 100%Connected Caveman Graph · network-topology · 95%Collective Predictive Coding · hypothesis · 90%

Relation Signals (3)

Active Inference → underpins → Multi-agent simulation model

confidence 100% · We propose a multi-agent simulation model grounded in active inference

Agents → interacton → Connected Caveman Graph

confidence 95% · The agents described above are situated on a social network... The experiments were conducted on a connected caveman graph

Metropolis–Hastings Naming Game → realizes → Collective Predictive Coding

confidence 85% · The Metropolis–Hastings Naming Game (MHNG) has been proposed as its concrete algorithmic realization

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Social agents both internalize collective norms and reshape them through creative action, yet computational models have not captured this bidirectional process within a unified framework. We propose a multi-agent simulation model grounded in active inference that formalizes the dialectical constitution of social reality on a structured social network. Each agent maintains an internal generative model, communicates with neighbors to form social priors, creates novel observations, and selectively incorporates others' creations into memory. Simulation experiments demonstrate three main findings. First, informationally cohesive social groups emerge endogenously, with representational alignment mirroring the cluster topology of the underlying network. Second, a circular mutual constitution arises between social representations and the observation distribution, maintained through agents' creative acts that project representational structure onto the external world. Third, the propagation of creations exhibits selective, heterogeneous patterns distinct from the stable diffusion of social representations, indicating that agents construct cultural niches through local interaction dynamics. These results suggest that the interplay between social conformity and creative deviation can give rise to the endogenous formation and differentiation of shared social reality.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

46,566 characters extracted from source content.

Expand or collapse full text

Social Reality Construction via Active Inference: Modeling the Dialectic of Conformity and Creativity Kentaro Nomura1,†, Takato Horii1,2 1The University of Osaka, Japan, 2IRCN, The University of Tokyo, Japan † nomura.kentaro.7ks@ecs.osaka-u.ac.jp Abstract Social agents both internalize collective norms and reshape them through creative action, yet computational models have not captured this bidirectional process within a unified framework. We propose a multi-agent simulation model grounded in active inference that formalizes the dialectical constitution of social reality on a structured social network. Each agent maintains an internal generative model, communicates with neighbors to form social priors, creates novel observations, and selectively incorporates others’ creations into memory. Simulation experiments demonstrate three main findings. First, informationally cohesive social groups emerge endogenously, with representational alignment mirroring the cluster topology of the underlying network. Second, a circular mutual constitution arises between social representations and the observation distribution, maintained through agents’ creative acts that project representational structure onto the external world. Third, the propagation of creations exhibits selective, heterogeneous patterns distinct from the stable diffusion of social representations, indicating that agents construct cultural niches through local interaction dynamics. These results suggest that the interplay between social conformity and creative deviation can give rise to the endogenous formation and differentiation of shared social reality. Submission type: Full Paper Data/Code available at: https://github.com/jemand-rkn/social-reality-aif.git Introduction Social agents both internalize collective norms and reshape them by creating and reifying new objects of meaning in the physical world. Through local interaction, these individual acts give rise to a shared social reality that, in turn, feeds back to shape further individual action. Mead (Mead, 1972) located this tension within the individual, formulating it as the interaction between the norm-internalizing Me and the creatively responding I. Berger and Luckmann (Berger and Luckmann, 1966), in contrast, characterized the same dynamic at the macro level as a dialectic of externalization, objectivation, and internalization, describing how society reproduces itself through individual action. Despite this rich theoretical tradition, the computational mechanisms by which bottom-up creative formation and top-down social regulation jointly produce stable yet evolving norms remain poorly understood. To address this gap, we propose a multi-agent simulation model grounded in the free energy principle. In this model, agents actively create novel observations and communicate via a naming game on a social network, thereby computationally realizing the dialectical process of social reality construction. Existing models in the Artificial Life community have demonstrated that shared conventions can emerge from local agent interactions, yet they uniformly treat agents as rule-following entities rather than probabilistic reasoners with internal generative models. The naming game paradigm (Steels, 1995; Baronchelli et al., 2006) established that shared vocabularies self-organize through pairwise interaction, undergoing a sharp phase transition to global consensus without centralized coordination. Similarly, the iterated learning framework (Smith et al., 2003; Kirby et al., 2008) showed that cultural transmission alone can generate compositional linguistic structure. Communicative exploration has also been shown to sustain open-ended collective novelty (Witkowski and Ikegami, 2019). However, because agents in these models lack internal generative models, they cannot capture the top-down regulatory influence of social context on individual cognition or the generative capacity to produce observations that deviate from prevailing norms. Active inference (Parr et al., 2022) offers a principled framework for deriving collective behavior and social norm formation from individual surprise minimization without requiring explicit behavioral rules. Canonical collective phenomena—cohesion, milling, and directed motion—have been shown to emerge when agents minimize surprise about their conspecifics, with real-time model updates coupling individual learning to group-level properties (Heins et al., 2024). Communication under active inference has further been shown to drive generalized synchrony and long-term convergence of generative models between agents (Friston and Frith, 2015a, b). Norm acquisition has also been formalized as free energy minimization with respect to a model of others’ expectations (Veissière et al., 2019; Kaufmann et al., 2021). However, these models address either spatial coordination or the internalization of pre-existing expectations; none accounts for how agents’ own creative acts feed back into the formation of shared social representations. This gap corresponds to the missing bottom-up direction in niche construction theory (Odling-Smee et al., 2003; Laland et al., 2000; Laland and O’Brien, 2011), which holds that organisms actively reshape their environment. This process has been formalized as free energy minimization through environmental modification, providing a variational basis for niche construction (Constant et al., 2018). The Collective Predictive Coding (CPC) hypothesis (Taniguchi, 2024) extends predictive coding and the free energy principle to the societal scale, offering a theoretical account of how shared representations such as symbol systems and social norms emerge through communicative interaction. Under CPC, social representations are formed through decentralized Bayesian inference. The Metropolis–Hastings Naming Game (MHNG) has been proposed as its concrete algorithmic realization (Hagiwara et al., 2019; Taniguchi et al., 2023), which has been further extended to N-agent populations through a recursive formulation (Inukai et al., 2023). However, CPC treats social representations as external shared variables that connect all agents through a joint generative model. This formulation implicitly assumes that a common representation is shared across the population, thereby precluding the natural emergence of distinct social groups. Furthermore, agents in this framework are passive observers that infer representations from observations they receive, leaving the feedback pathway from individual creative acts to social representation formation unaddressed. In the present work, social representations are instead placed inside each agent and allowed to differ across individuals, enabling representational alignment to emerge naturally from local interaction. Agents are also endowed with the capacity for active creation, closing the feedback loop between individual creativity and collective norm formation. We present a multi-agent simulation model that formalizes the bidirectional constitution of social reality within a unified active inference framework on a structured social network. Each agent is formalized as an internal generative model that is continuously updated through communication with neighbors, active creation of novel observations, and selective incorporation of others’ creations. Top-down regulation and bottom-up formation are realized as variational and expected free energy minimization, respectively, placing them in a formally adversarial yet complementary relationship. The present work makes three contributions. First, we show that informationally cohesive social groups emerge endogenously, mirroring the cluster topology of the underlying network. Second, social representations and the observation distribution are mutually constituted through agents’ creative acts. Third, creations propagate in selective, heterogeneous patterns, indicating that agents construct cultural niches through local interaction. Proposed Model Agent Architecture Each agent in the proposed model maintains an internal generative model that describes how observations are generated from social representations as latent variables. Here, observations represent information that agents experience in the physical world. The generative model is agent-specific and is updated in a fully decentralized manner. Specifically, the generative model for agent AkA^k is defined as follows: social prior: ∼psocialk(), z p^k_ social( z), (1) likelihood: ∼pθk(|), o p^k_θ( o| z), (2) posterior: ∼qϕk(|), z q^k_φ( z| o), (3) where z and o denote the social representation and the observation, respectively, and k indexes the agent. The likelihood pθk(|)p^k_θ( o| z) and the approximate posterior qϕk(|)q^k_φ( z| o) are parameterized by neural network weights θ and ϕφ, respectively. The social prior psocialk(⋅)p^k_ social(·) is a probability distribution that reflects the social context in which the agent is situated. This distribution evolves as the agent interacts with others. Each agent also possesses a discriminator Dψk(,)D^k_ψ( z, o) that quantifies the divergence between the agent’s approximate posterior qϕk(⋅|)q^k_φ(·| o) and the social prior psocialk(⋅)p^k_ social(·). The discriminator is trained to approximate the log-density ratio between these two distributions: Dψk(,)≈log⁡qϕk(|)psocialk().D^k_ψ( z, o)≈ q^k_φ( z| o)p^k_ social( z). (4) The discriminator enables each agent to assess how much its own observation-conditioned beliefs deviate from the social expectations formed through interaction with its neighbors. This discriminator plays a central role in both model updating and the creation of novel observations. Each agent further maintains a memory buffer ℬkB^k that stores artifacts as a first-in-first-out (FIFO) data structure. Newly created artifacts are appended to the buffer, while the oldest entries are discarded. In addition to the agent’s own creations, artifacts produced by other agents may be incorporated into ℬkB^k through selective replacement. When the agent draws samples from ℬkB^k for model updating or communication, the stored artifacts serve as observations for its generative model. Creation of Novel Artifacts via Active Inference Each agent creates novel artifacts + o_+ that minimize the expected free energy within the framework of active inference. The expected free energy kG^k for agent AkA^k is defined as: k(+) ^k( o_+) =−DKL[qϕk(|+)∥psocialk()]−qϕk[log⁡pθk(+|)] =-D_KL[q^k_φ( z| o_+)\|p^k_ social( z)]-E_q^k_φ[ p^k_θ( o_+| z)] (5) ≈−qϕk[Dψk(,+)]−qϕk[log⁡pθk(+|)]. ≈-E_q^k_φ[D^k_ψ( z, o_+)]-E_q^k_φ[ p^k_θ( o_+| z)]. In practice, the second term is weighted by a coefficient λ. The optimization is performed via gradient descent, initialized from an artifact sampled from the agent’s memory ℬkB^k, and the resulting artifact is subsequently appended to ℬkB^k. The two terms in Eq. (5) jointly drive agents to explore artifacts that are novel relative to the prevailing social context while remaining consistent with their own generative models. The first term evaluates the negative divergence between the approximate posterior conditioned on the created artifact + o_+ and the social prior. Minimizing this term encourages the agent to produce creations that deviate from the established social prior, thereby maximizing epistemic information gain regarding social representations. The second term corresponds to the reconstruction error of + o_+ under the agent’s generative model, which penalizes excessive deviation from the agent’s individual knowledge. Through simultaneous optimization of both terms, each agent produces artifacts that represent physical entities not yet captured by the prevailing social context, while preserving the epistemological consistency of its own model. Note that, we formulate this process as directly sampling observations that minimize the expected free energy, without introducing explicit actions, thereby simplifying the model. Model Update The discriminator Dψk(,)D^k_ψ( z, o) is trained under the f-GAN framework (Nowozin et al., 2016), which generalizes generative adversarial networks (Goodfellow et al., 2014) to f-divergences for implicitly learning unknown probability distributions through adversarial training. Specifically, we adopt the reverse Kullback–Leibler divergence as the f-divergence to estimate the divergence between the social prior and the approximate posterior. Let kP^k denote the set of observation–representation pairs sampled from the social prior through inter-agent communication. The loss function for training DψkD^k_ψ is: ℒk(ψ)=(,)∼k[exp⁡(Dψk(,))]−∼ℬk,qϕk(|)[Dψk(,)]. splitL^k(ψ)=&E_( o, z) ^k[ (D^k_ψ( z, o))]\\ &-E_ o ^k,\,q^k_φ( z| o)[D^k_ψ( z, o)]. split (6) The generative model parameters θ and ϕφ are updated by minimizing the variational free energy ℱkF^k, which is expressed using the discriminator as: ℱk(θ,ϕ) ^k(θ,φ) =DKL[qϕk(|)∥psocialk()]−qϕk[log⁡pθk(|)] =D_KL[q^k_φ( z| o)\|p^k_ social( z)]-E_q^k_φ[ p^k_θ( o| z)] (7) ≈qϕk[Dψk(,)]−qϕk[log⁡pθk(|)]. _q^k_φ[D^k_ψ( z, o)]-E_q^k_φ[ p^k_θ( o| z)]. In practice, the first term is weighted by a coefficient β. Because Eq. (7) contains the positive value of DψkD^k_ψ in the first term, minimizing the variational free energy drives each agent’s approximate posterior toward the social prior, thereby aligning the agent’s internal model with the prevailing social context. The model update and the creation process stand in an adversarial relationship. Comparing Eq. (7) with Eq. (5) reveals that the sign of the term involving DψkD^k_ψ is reversed between the two objectives. Variational free energy minimization during model updating pushes the approximate posterior toward the social prior, whereas expected free energy minimization during creation and selective memorization pushes it away. The former thus constitutes a top-down regulatory influence from the social context, while the latter represents a bottom-up contribution to social formation through exploratory creative acts. Social Network The agents described above are situated on a social network that governs their communication topology and determines the scope of social interaction. The network is formalized as an undirected graph (,ℰ)(K,E), where =A1,…,AKK=\A^1,…,A^K\ denotes the set of K agents and ℰE denotes the set of edges. At each time step, agent AkA^k exchanges inferred social representations and creations with its neighboring agents Ak′∈(k)A^k (k), where (k)N(k) is the set of agents connected to AkA^k by an edge. Through these pairwise exchanges, each agent’s social prior and memory are shaped not only by its own internal processes but also by the creations and representations of its neighbors, as described in the following subsections. Selective Memorization of Others’ Creations Whether to accept a creation received from another agent and update one’s memory can be understood as action selection under active inference. In active inference, an agent selects actions a according to a Boltzmann distribution over the expected free energy: P(a)∝exp⁡(−k(a)/τ)P(a) (-G^k(a)/τ), where τ is a temperature parameter controlling the stochasticity of the selection (Parr et al., 2022). We consider two candidate actions for each received creation: acceptance (a=+′a= o _+) and rejection (a=−a= o_-). Here, +′ o _+ denotes a creation received from a neighboring agent in (k)N(k), and −∼ℬk o_- ^k is an observation sampled from the agent’s own memory. Evaluating kG^k for both candidates and comparing their energy values yields the acceptance probability: ro=min⁡(1,exp⁡(−k(+′)−k(−)τ)),r_o= (1, (- G^k( o _+)-G^k( o_-)τ ) ), (8) and upon acceptance, the memory buffer is updated as: ℬk←(ℬk∖−)∪+′.B^k←(B^k \ o_-\)∪\ o _+\. (9) In the limit τ→0τ→ 0, the agent deterministically selects the creation with lower expected free energy. This pairwise comparison enables each agent to judge whether another agent’s creation better aligns with its current interests than its own past creations, and to selectively retain the more relevant artifact in memory. Communication via Social Representations Agents construct their social priors by exchanging social representations with neighbors on the social network and integrating the received representations with their own models. Through this communicative process, each agent forms prior knowledge that reflects the perspective of a generalized social member as perceived from its own viewpoint. We assume joint attention during communication: both communicating agents present their memorized artifacts and creations, and infer social representations conditioned on these shared artifacts. To obtain samples from each agent’s social prior psocialk()p^k_ social( z) under joint attention, we employ the Metropolis–Hastings Naming Game (MHNG) (Taniguchi et al., 2023). MHNG is a method for sampling latent variables that can be shared between two agents who jointly attend to the same artifacts. Although MHNG originally yields samples from an approximate posterior qk()q^k( z), we identify these with the social prior via Bayesian updating, i.e., psocialk()=qk()p^k_ social( z)=q^k( z). Let k,k′=k∪k∪k′∪k′O^k,k =C^k ^k ^k ^k denote the union of creations ⋆C and memory samples ⋆⊂ℬ⋆D from agent AkA^k and its neighbor Ak′∈(k)A^k (k). Notably, not only the creations ⋆C but also past memories ⋆D serve as objects of joint attention. Both agents infer social representations z from observations in k,k′O^k,k using their respective inference models qϕ⋆q _φ. Let ⋆∼qϕ⋆(⋅|) z q _φ(·| o) denote the social representation inferred by agent A⋆A from ∼k,k′ o ^k,k . Agent AkA^k accepts the representation k′ z^k received from its neighbor with probability: rz=min⁡(1,pθk(|k′)pθk(|k)),r_z= (1, p^k_θ( o| z^k )p^k_θ( o| z^k) ), (10) and upon rejection retains its own inferred representation k z^k as a sample from psocialk()p^k_ social( z). The set of observation–representation pairs kP^k used for discriminator training is constructed through this communication process. Each pair in kP^k consists of an observation drawn from ⋃k′=♯(A∈(k))k,k′ _k = (A (k))O^k,k and the corresponding social representation accepted or retained through the Metropolis–Hastings procedure above. Simulation Procedure Algorithm 1 Simulation procedure of the proposed model 1: for each time step t=1,2,…t=1,2,… do 2: // Creation 3: for each agent Ak∈A^k do 4: Create +←arg⁡min⁡k() o_+← _ oG^k( o); append to ℬkB^k 5: end for 6: // Selective memorization 7: for each agent Ak∈A^k do 8: Receive +′ o _+ from Ak′∈(k)A^k (k); accept into ℬkB^k with prob. ror_o 9: end for 10: // Communication via MHNG 11: for each edge (Ak,Ak′)∈ℰ(A^k,A^k ) do 12: Exchange representations k,k′ z^k, z^k over joint observations k,k′O^k,k 13: Accept k′ z^k with prob. rzr_z; collect pairs into kP^k 14: end for 15: // Model update 16: for each agent Ak∈A^k do 17: Update ψkψ^k by minimizing ℒk(ψ)L^k(ψ); update θk,ϕkθ^k,φ^k by minimizing ℱk(θ,ϕ)F^k(θ,φ) 18: end for 19: end for The proposed model iterates through four stages at each time step, cycling between creative exploration and social conformity. Algorithm 1 summarizes the overall procedure. First, each agent infers a novel artifact + o_+ by minimizing the expected free energy kG^k and appends it to its memory ℬkB^k. Second, creations received from neighboring agents are selectively incorporated into ℬkB^k under active inference. Third, each agent executes MHNG with its neighbors to obtain samples from the social prior psocialk()p^k_ social( z). Finally, the discriminator and generative model parameters θ, ϕφ, and ψ are updated. Each stage fulfills a distinct functional role within the overall dynamics. The creation and selective memorization stages modify the agent’s memory ℬkB^k. The communication stage shapes the social prior. The model update stage adapts the generative model and discriminator to the evolving social context. Through the cyclic repetition of these stages, the formation of social representations and creative deviation mutually drive one another. Experiments We conducted a series of simulation experiments to verify that the proposed model gives rise to top-down regulatory effects from the collective society and bottom-up social formation through the creative acts of individual agents, and to observe what emergent phenomena result from these interactions. Experimental Settings Figure 1: The social network used in the experiments. A: Network structure (blue: cluster 0, orange: cluster 1). B: Adjacency matrix. The experiments were conducted on a connected caveman graph (Watts, 1999) comprising K=14K=14 agents. This social network topology consists of two cluster-structured subgraphs, each containing seven agents. We refer to these subgraphs as cluster 0 (agents k=0k=0–66) and cluster 1 (agents k=7k=7–1313). Figure 1 shows the network topology and its corresponding adjacency matrix. Observations o were represented as two-dimensional continuous real-valued vectors, and social representations z as four-dimensional continuous real-valued vectors. Each agent’s memory buffer ℬkB^k had a capacity of 7,000 samples. The buffer was initialized with samples drawn from Gaussian distributions whose peaks were placed at distinct locations in observation space for each agent (see the leftmost panel of Fig. 4). The generative model parameters θ and ϕφ of each agent were initialized by pre-training on the initial memory ℬkB^k within the variational autoencoder (Kingma and Welling, 2014) framework. The discriminator parameters ψ were initialized randomly. The hyperparameters were set as follows. The weighting coefficients for the second term of Eq. (5) and the first term of Eq. (7) were set to λ=0.1λ=0.1 and β=1.0β=1.0, respectively. The temperature parameter of the Boltzmann distribution used in selective memorization was set to τ=0.3τ=0.3. The simulation was run for 5,000 steps. At each step, each agent created |k|=6|C^k|=6 novel observations, and the number of samples drawn from memory ℬkB^k for communication was |k|=100|D^k|=100. Model parameters were updated for 5 iterations per step with a learning rate of 1×10−51× 10^-5, using the Adam optimizer for both the creation and model-update phases. The mini-batch size for sampling from ℬkB^k during model updates was set to 256. Experimental Conditions To assess the impact of the creation and selective memorization processes on the emergent social representations and their relationship to creations, we compared the following two experimental conditions: w/ creation Agents perform creation and selective memorization. The simulation follows the full procedure. w/o creation Agents perform neither creation nor selective memorization. During the communication phase, the set of creative artifacts shared in the MHNG is set to k=∅C^k= . Results Temporal Evolution of Social Representations Figure 2: Evolution of the average Wasserstein distance matrix computed between the social representations inferred by agents. Figure 3: MDS embedding trajectory of the GW distance matrix computed from the inferred social representation structures. Numbers in black indicate time steps. We characterized the relational structure of social representations by comparing the representations inferred by each agent in the w/ creation condition. We first constructed a reference observation set refO_ref by uniformly sampling observations from each agent’s memory at each time step. Each agent’s inference model was then applied to every element of refO_ref, yielding the posterior distribution qϕk(⋅|)q^k_φ(·| o) over social representations for all agents. For every agent pair (k,k′)(k,k ), we computed the mean Wasserstein distance between the per-observation posteriors as a measure of how similarly two agents represent the same observation. The Wasserstein distance matrix reveals a gradual transition from globally uniform inter-agent distances to a block-diagonal structure (Fig. 2). Initially, representational distances were roughly uniform across all agent pairs. As the simulation progressed, intra-cluster distances decreased relative to inter-cluster distances, indicating that representational alignment mirrored the cluster structure of the social network. As a result, informationally cohesive social groups—composed of frequently communicating individuals—emerged endogenously from the dynamics of the model. However, Wasserstein distance captures only point-wise dissimilarity and does not reflect whether the relational geometry among representations is shared across agents, even when absolute values differ. To evaluate this structural similarity, we additionally computed the Gromov–Wasserstein (GW) distance (Mémoli, 2011), which quantifies the degree to which the relational structure of an entire set of representations is preserved across agents. The GW distance, visualized through multidimensional scaling (MDS) embedding, reveals a systematic, cluster-aligned divergence in representational structure over time (Fig. 3). Initially, all agents were densely clustered near the origin, indicating a shared relational structure. Over the course of the simulation, agents in cluster 0 migrated toward the upper region of the embedding space, while those in cluster 1 moved downward, reflecting a divergence that mirrors the community boundaries of the social network. Hub agents exhibited a distinct trajectory that reflected their simultaneous membership in both communities. Agents k=0,7k=0,7, and 1313, which bridge the two clusters through direct inter-cluster connections, did not converge toward either cluster’s trajectory in the MDS embedding (Fig. 3) but instead remained in the intermediate region, oscillating between the two groups throughout the simulation. This behavior is consistent with the expectation that agents exposed to competing social priors from both clusters cannot settle into a single cluster-specific representational structure. Representational differentiation proceeded in two stages: agents first diverged in the values assigned to individual observations, and only later diverged in the relational structure among those values. The Wasserstein distance matrix shows clear block-diagonal structure by time steps 1001–2000 (Fig. 2), indicating that agents in different clusters had already developed distinct per-observation representations. In contrast, the MDS embedding of the GW distance matrix shows that agents from both clusters remained tightly overlapping until approximately step 2000, with cluster-level separation emerging only thereafter (Fig. 3). This temporal lag suggests that local differences in created observations must first accumulate sufficiently before they reshape the global geometry of each agent’s representational space. Temporal Evolution of Memorized Observations Figure 4: Temporal evolution of the distribution of observations created and memorized by each agent. The observations created and memorized by each agent underwent a characteristic temporal trajectory: initial convergence aligned with the cluster structure, followed by the emergence of individually distinctive observations within each cluster. Figure 4 shows the temporal change in the distribution of observations under the w/ creation condition. At the start of the simulation, each agent’s memorized observations were distributed across distinct positions in the observation space. From the onset through step 1500, these observations gathered into arch-shaped structures on a cluster-by-cluster basis, with each cluster developing a characteristic distribution and internal geometry. Following this cluster-level consolidation, individual agents began to construct genuinely novel observations beyond the existing observation region. Notably, agents such as agent 6 and agent 13 were among the first to expand the occupied region of the observation space. Toward the end of the simulation, the coherent cluster-level structures began to dissolve, and individual agents progressively created and memorized observations unique to themselves. At this stage, a small number of agents came to occupy nearby positions in the observation space, suggesting the emergence of finer-grained affinities that cut across the original cluster boundaries. This temporal trajectory can be interpreted as the outcome of an adversarial interplay between social conformity, induced by model updates in response to neighbors’ observations, and creative deviation, driven by each agent’s generative action under active inference. In the earlier phase, agents prioritized conforming to their immediate social context through mutual creation and memorization within clusters, gradually constructing a shared observation structure. Once this cluster-level alignment was sufficiently established, agents shifted toward creative exploration, maintaining individual distinctiveness while remaining loosely connected to their cluster’s shared structure. This dynamic reflects how each agent actively constructed its external world under the dual influence of creative curiosity and social pressure. Structural Similarity Between Observations and Social Representations Figure 5: RSA similarity between observations and social representations for each cluster, smoothed with a 25-step moving average. Shaded bands indicate standard deviation. To examine the effect of agents’ creative acts on social representation formation, we compared the structural similarity between memorized observations and inferred social representations under the w/ creation and w/o creation conditions. Simulations were conducted with five different random seeds for each condition. For each cluster, we quantified the similarity using representational similarity analysis (RSA) (Kriegeskorte et al., 2008) applied to the Euclidean distance matrices of the memorized observations and the social representations. Figure 5 shows the temporal evolution of this similarity under both conditions. Agents’ creative acts were essential for maintaining structural alignment between observations and social representations. Under the w/o creation condition, similarity decreased monotonically after the 500th step in both clusters. In contrast, under the w/ creation condition, similarity remained at approximately 0.65 or above throughout the simulation. This finding implies that creative acts function as a mechanism through which the structure of social representations is projected onto the external observation space. That is, the regulatory influence of social representations extended beyond the agents to shape the external world, with individual agents serving as the mediating mechanism. These results indicate that the top-down regulatory influence of social representations and the bottom-up creative acts of individual agents operate in a mutually reinforcing cycle, through which social representations and the observation distribution are co-constituted over time. Propagation of Social Representations and Creations Figure 6: Average acceptance rate of social representations and creations between agents, aggregated over successive intervals of time steps. Edge transparency reflects the frequency with which communication resulted in a zero acceptance rate. We analyzed the acceptance rates of social representations and creations across agent pairs to investigate their propagation dynamics (Fig. 6). Edge transparency reflects the frequency with which communication resulted in a zero acceptance rate; edges become more transparent as the proportion of zero-acceptance time steps increases. The stable intra-cluster propagation of social representations accounts for the sustained representational alignment observed across the preceding analyses (Figs. 2 and 3). Within each cluster, edges were rarely pruned, and social representations were consistently accepted in a reciprocal manner throughout the simulation. This persistent mutual acceptance generated continuous pressure for each agent to internalize fellow cluster members’ representations, anchoring individual generative models to the shared cluster structure. Meanwhile, inter-cluster edges underwent pruning in the latter half of the simulation, indicating that social representations became increasingly difficult to share across cluster boundaries. In contrast, the propagation pathways of creations shifted dynamically over time, accounting for the transient inter-agent similarities observed in the observation space (Fig. 4). Within clusters, edge pruning occurred such that creations came to be shared only among a small subset of agent pairs, and these pruning patterns shifted over a large timescale. Across clusters, creations were consistently accepted between the hub agents, agent 0 and agent 13. Furthermore, in the latter half of the simulation, agent 0 frequently shared creations with agents in cluster 1 while ceasing to share with agents in its own cluster 0. The observation that a small number of agents came to produce similar creations at any given time can thus be attributed to these dynamically reconfiguring acceptance pathways. The temporal shift in such agent combinations further reflects the ongoing reconfiguration of acceptance patterns. For instance, the sustained convergence between agent 0 and agent 13 corresponds to their persistent mutual acceptance of creations across cluster boundaries. Discussion The simulation results demonstrate that the model gives rise to informationally coherent social groups whose representational structure mirrors the cluster topology of the underlying network. Agents within the same cluster converged toward similar generative models, while inter-cluster divergence became progressively more pronounced. Moreover, in the w/ creation condition, a circular mutual constitution emerged: agents’ generative models shaped the direction of their creative outputs, which propagated through the network and in turn reshaped the social representations of other agents. The persistent structural alignment between representational and observational spaces confirms that creative action is functionally integral to the coherence of social reality, not merely a byproduct of individual expression. The analysis of propagation dynamics further reveals a dual mechanism that sustains social representations amid continuously changing observations. Within each cluster, sustained mutual acceptance of representations through the naming game generates a persistent internalization pressure that draws individual models toward a shared attractor. This pressure maintains collective coherence even as the observational landscape is continually altered by creative action. In contrast, the reception of creative artifacts became progressively concentrated among specific agent pairs, indicating that agents effectively constructed cultural niches. This selective uptake points to a process of cultural differentiation driven not by top-down prescription but by the self-organizing dynamics of local interaction. The two core processes of the model—generative model update and active inference-driven creation—appear to operate in opposition but are in fact complementary. Model update aligns agents’ generative models with the prevailing representational consensus by internalizing social context and norms. Creative action, by contrast, maintains epistemic homeostasis while generating observations that deviate from the existing social context, thereby diversifying the distribution of observations available for social exchange. Conversely, the conformity achieved through model update provides a stable representational basis from which further creative deviation can proceed. This complementary relationship between conformity and deviation constitutes the core mechanism driving the process of social reality construction modeled here. The present model has several limitations that suggest directions for future work. First, the social network is fixed throughout the simulation, precluding both active partner selection by agents and the self-organized emergence of communication topology. Introducing spatial position and movement as variables subject to active inference would allow the network structure to co-evolve with agents’ generative models. Second, the model abstracts away embodiment, such that agents generate observations without mediation by a body or physical environment. Incorporating action variables operating through a body–environment system would enable modeling the full causal chain from creative intention to observable artifact. Beyond these extensions, information-theoretic tools such as non-trivial information closure (NTIC) (Chang et al., 2020; Dobata et al., 2025) offer a complementary analytical direction. Applying NTIC to quantify the degree of autonomy in each agent’s behavior would provide a formal measure of social autonomy, contributing to a mathematical understanding of the emergence of social self as discussed by Mead. Conclusion This study presents a computational model of social reality construction that integrates the regulatory influence of social structure with individual creative action under the framework of active inference. Agents are formalized as generative models that are continuously updated through communication on a social network, capturing both the internalization of social norms and the generation of novel observations through creative action. This dual formulation allowed us to investigate the mechanisms by which social representations emerge, stabilize, and diverge over time. Acknowledgements This work was supported by JSPS grant JP23H04834. References A. Baronchelli, M. Felici, V. Loreto, E. Caglioti, and L. Steels (2006) Sharp transition towards shared vocabularies in multi-agent systems. J. Stat. Mech. 2006 (06), p. P06014–P06014 (en). Cited by: Introduction. P. L. Berger and T. Luckmann (1966) The social construction of reality: a treatise in the sociology of knowledge. [1st ed.] edition, Doubleday, Garden City, N.Y. (en). Cited by: Introduction. A. Y. C. Chang, M. Biehl, Y. Yu, and R. Kanai (2020) Information closure theory of consciousness. Front. Psychol. 11, p. 1504 (en). Cited by: Discussion. A. Constant, M. J. D. Ramstead, S. P. L. Veissière, J. O. Campbell, and K. J. Friston (2018) A variational approach to niche construction. J. R. Soc. Interface 15 (141), p. 20170685 (en). Cited by: Introduction. S. Dobata, Y. Notomi, T. Kudo, I. Horiguchi, M. Crosscombe, N. Maruyama, A. Y. Chang, and T. Ikegami (2025) Collective behavior of water beetles with varying degrees of information closure. In ALIFE 2025: Ciphers of Life: Proceedings of the Artificial Life Conference 2025, Vol. 37, p. 47 (en). Cited by: Discussion. K. Friston and C. Frith (2015a) A duet for one. Conscious. Cogn. 36, p. 390–405 (en). Cited by: Introduction. K. J. Friston and C. D. Frith (2015b) Active inference, communication and hermeneutics. Cortex 68, p. 129–143 (en). Cited by: Introduction. I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio (2014) Generative adversarial nets. Advances in Neural Information Processing Systems 27. Cited by: Model Update. Y. Hagiwara, H. Kobayashi, A. Taniguchi, and T. Taniguchi (2019) Symbol emergence as an interpersonal multimodal categorization. Front. Robot. AI 6, p. 134 (en). Cited by: Introduction. C. Heins, B. Millidge, L. Da Costa, R. P. Mann, K. J. Friston, and I. D. Couzin (2024) Collective behavior from surprise minimization. Proc. Natl. Acad. Sci. U. S. A. 121 (17), p. e2320239121 (en). Cited by: Introduction. J. Inukai, T. Taniguchi, A. Taniguchi, and Y. Hagiwara (2023) Recursive metropolis-hastings naming game: symbol emergence in a multi-agent system based on probabilistic generative models. Front. Artif. Intell. 6, p. 1229127 (en). Cited by: Introduction. R. Kaufmann, P. Gupta, and J. Taylor (2021) An active inference model of collective intelligence. Entropy (Basel) 23 (7), p. 830 (en). Cited by: Introduction. D. P. Kingma and M. Welling (2014) Auto-encoding variational bayes. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, Y. Bengio and Y. LeCun (Eds.), External Links: Link Cited by: Experimental Settings. S. Kirby, H. Cornish, and K. Smith (2008) Cumulative cultural evolution in the laboratory: an experimental approach to the origins of structure in human language. Proc. Natl. Acad. Sci. U. S. A. 105 (31), p. 10681–10686 (en). Cited by: Introduction. N. Kriegeskorte, M. Mur, and P. Bandettini (2008) Representational similarity analysis - connecting the branches of systems neuroscience. Front. Syst. Neurosci. 2, p. 4 (en). Cited by: Structural Similarity Between Observations and Social Representations. K. N. Laland, J. Odling-Smee, and M. W. Feldman (2000) Niche construction, biological evolution, and cultural change. Behav. Brain Sci. 23 (1), p. 131–46; discussion 146–75 (en). Cited by: Introduction. K. N. Laland and M. J. O’Brien (2011) Cultural niche construction: an introduction. Biol. Theory 6 (3), p. 191–202 (en). Cited by: Introduction. G.H. Mead (1972) Mind self and society: from the standpoint of a social behaviorist. Phoenix books, University of Chicago Press. External Links: Link Cited by: Introduction. F. Mémoli (2011) Gromov–Wasserstein distances and the metric approach to object matching. Found. Comput. Math. 11 (4), p. 417–487 (en). Cited by: Temporal Evolution of Social Representations. S. Nowozin, B. Cseke, and R. Tomioka (2016) F-GAN: training generative neural samplers using variational divergence minimization. Advances in Neural Information Processing Systems 29. Cited by: Model Update. F. J. Odling-Smee, K. N. Laland, and M. W. Feldman (2003) Niche construction: the neglected process in evolution (MPB-37). Monographs in Population Biology, Princeton University Press, Princeton, NJ (en). Cited by: Introduction. T. Parr, G. Pezzulo, and K. J. Friston (2022) Active inference: the free energy principle in mind, brain, and behavior. The MIT Press. External Links: ISBN 9780262369978, Document, Link, https://direct.mit.edu/book-pdf/2246566/book_9780262369978.pdf Cited by: Introduction, Selective Memorization of Others’ Creations. K. Smith, S. Kirby, and H. Brighton (2003) Iterated learning: a framework for the emergence of language. Artif. Life 9 (4), p. 371–386 (en). Cited by: Introduction. L. Steels (1995) A self-organizing spatial vocabulary. Artif. Life 2 (3), p. 319–332 (en). Cited by: Introduction. T. Taniguchi, Y. Yoshida, Y. Matsui, N. Le Hoang, A. Taniguchi, and Y. Hagiwara (2023) Emergent communication through metropolis-hastings naming game with deep generative models. Adv. Robot. 37 (19), p. 1266–1282 (en). Cited by: Introduction, Communication via Social Representations. T. Taniguchi (2024) Collective predictive coding hypothesis: symbol emergence as decentralized bayesian inference. Front. Robot. AI 11, p. 1353870 (en). Cited by: Introduction. S. P. L. Veissière, A. Constant, M. J. D. Ramstead, K. J. Friston, and L. J. Kirmayer (2019) Thinking through other minds: a variational approach to cognition and culture. Behav. Brain Sci. 43 (e90), p. e90 (en). Cited by: Introduction. D. J. Watts (1999) Networks, dynamics, and the small‐world phenomenon. Am. J. Sociol. 105 (2), p. 493–527 (en). Cited by: Experimental Settings. O. Witkowski and T. Ikegami (2019) How to make swarms open-ended? evolving collective intelligence through a constricted exploration of adjacent possibles. Artif. Life 25 (2), p. 178–197 (en). Cited by: Introduction.

Instant research discovery