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

This paper proposes a multi-agent active inference model demonstrating how the dialectical interplay between conformity and creativity on structured networks endogenously generates shared social realities while simultaneously enabling the formation of distinct cultural niches through selective creative propagation.

The Gist

Imagine a bustling city where everyone is trying to figure out what "reality" looks like, but no one has a map. Some people want to stick to the rules everyone else follows (conformity), while others want to paint the walls with new, weird colors (creativity).

This paper by Kentaro Nomura and Takato Horii is like a computer simulation of that city. They built a group of digital "agents" (think of them as tiny, invisible robots) to see how a shared reality forms when these two forces—fitting in and standing out—fight and dance together.

Here is the story of what they found, explained simply:

The Setup: The Robots and the Rules

The researchers put these robots on a social network (like a digital version of a neighborhood). Each robot has two main jobs:

1. To Learn: They listen to their neighbors to understand what the group thinks is "normal."
2. To Create: They try to make up something new and interesting to show the group.

The magic happens because the robots use a special logic called Active Inference. You can think of this as a robot's internal compass. It constantly asks: "Does what I see match what I expect? If not, should I change my mind to fit in, or should I change the world to match my new idea?"

The Experiment: Two Clusters

They set up the robots in two distinct groups (let's call them Team Blue and Team Orange).

• Team Blue talks mostly to other Blue robots.
• Team Orange talks mostly to other Orange robots.
• There are a few "bridge" robots that talk to both teams.

They ran the simulation for a long time to see what happened.

The Three Big Discoveries

1. The "Echo Chambers" Form Naturally

At first, everyone was a bit confused and sounded the same. But as time went on, Team Blue started thinking in a very specific way, and Team Orange started thinking in a totally different way.

• The Analogy: Imagine two groups of friends at a party. One group starts a joke that only they understand, and they keep telling it until it becomes their inside language. The other group does the same with a different joke. Eventually, the two groups speak such different "languages" that they can barely understand each other anymore.
• The Result: The robots naturally formed tight-knit groups that mirrored the structure of their network. They didn't need a boss to tell them to separate; they just did it by talking to their neighbors.

2. The "Self-Fulfilling Prophecy" of Creativity

This is the most fascinating part. The robots didn't just observe the world; they made the world.

• The Analogy: Imagine a group of artists. If they all decide that "blue is the color of sadness," they start painting everything blue. Because they keep painting blue, the world becomes blue. Their creativity actually changed the reality they live in.
• The Result: The robots created new "observations" (new ideas or objects) that matched their group's beliefs. By creating these things, they reinforced their own beliefs. It's a loop: Beliefs create reality, and that reality strengthens beliefs. Without the robots creating new things, their shared reality would have crumbled.

3. The "Cultural Niches" (Who Shares What?)

The researchers noticed something cool about how ideas spread.

• The Analogy: Think of "Social Norms" (like "we wear shoes") as a slow-moving river. It flows steadily through the whole group. But "New Creations" (like a weird new dance move) are like fireflies. They don't flow in a river; they jump from person to person in unpredictable patterns.
• The Result:
  • Norms stayed stable within the groups (Blue stayed Blue).
  • New Creations jumped around wildly. Sometimes a robot in Team Blue would share a cool new idea with a robot in Team Orange, and they would both love it, even if their groups were fighting.
  • This created "cultural niches"—small pockets of connection that crossed the big group lines. It showed that while we might belong to a big tribe, our creativity allows us to make friends with people outside our tribe.

The Big Picture: The Tug-of-War

The paper concludes that society isn't just about following rules or just about being wild and free. It's a tug-of-war between the two.

• Conformity (fitting in) keeps the group stable so everyone can understand each other.
• Creativity (standing out) keeps the group fresh and prevents it from getting stuck.

If you only have conformity, society becomes a boring, stagnant prison. If you only have creativity, society is chaotic noise where no one understands anyone. The "Social Reality" we live in is the beautiful, messy balance point where we agree on enough to communicate, but create enough to keep things interesting.

Why Does This Matter?

This model helps us understand how human culture works. It suggests that we are not just passive receivers of culture. We are active builders. Every time you post a new meme, start a new trend, or even just have a unique opinion, you are slightly reshaping the "social reality" of your community. We are all architects of the world we live in, constantly negotiating between "what everyone else does" and "what I want to do."

Technical Summary

1. Problem Statement

The paper addresses a fundamental gap in computational social science: the lack of a unified framework that captures the bidirectional process of social reality construction. Specifically, existing models fail to simultaneously account for:

• Top-down regulation: How agents internalize collective norms and social contexts.
• Bottom-up formation: How agents creatively reshape these norms through novel actions and observations.

Traditional models (e.g., Naming Games, Iterated Learning) often treat agents as rule-following entities or passive observers, lacking internal generative models. Consequently, they cannot explain how individual creative deviations feed back to alter shared social representations, nor how social structures emerge endogenously from local interactions without centralized coordination.

2. Methodology

The authors propose a multi-agent simulation model grounded in the Free Energy Principle and Active Inference. The system operates on a structured social network where agents act as probabilistic reasoners with internal generative models.

A. Agent Architecture

Each agent $A_k$ maintains:

1. Internal Generative Model: Describes how observations ($o$) are generated from latent social representations ($z$).
   • Social Prior ($p_{social}$): A distribution reflecting the social context, updated via interaction.
   • Likelihood ($p_\theta(o|z)$) & Posterior ($q_\phi(z|o)$): Parameterized by neural networks.
2. Discriminator ($D_\psi$): Trained via an f-GAN framework to approximate the log-density ratio between the agent's posterior and the social prior. This quantifies how much an agent's beliefs deviate from social expectations.
3. Memory Buffer ($B_k$): A FIFO structure storing artifacts (observations) created by the agent or accepted from neighbors.

B. Core Mechanisms

The simulation iterates through four stages per time step, balancing conformity and creativity:

1. Creative Exploration (Active Inference):

   • Agents generate novel artifacts ($o^+$) by minimizing Expected Free Energy ($G_k$).
   • $G_k$ consists of two competing terms:
     • Epistemic Value: Minimizing divergence from the social prior (encouraging deviation/novelty).
     • Pragmatic Value: Minimizing reconstruction error (ensuring the creation remains consistent with the agent's internal model).
   • This process allows agents to create observations that are novel relative to the current social context but epistemologically consistent.

2. Selective Memorization:

   • Agents receive creations from neighbors and decide whether to accept them into memory based on Active Inference action selection.
   • Acceptance is probabilistic (Boltzmann distribution) based on the expected free energy of the new artifact versus the agent's current memory samples. This allows agents to selectively retain "better" or more relevant cultural artifacts.

3. Communication (Social Prior Formation):

   • Agents engage in a Metropolis–Hastings Naming Game (MHNG) with neighbors.
   • They jointly attend to shared artifacts (creations and memories) to infer social representations ($z$).
   • Representations are exchanged and accepted/rejected based on likelihood ratios, effectively updating the Social Prior ($p_{social}$) for each agent.

4. Model Update:

   • Discriminator Training: Updates $\psi$ to better distinguish between social priors and individual posteriors.
   • Generative Model Update: Agents minimize Variational Free Energy ($F_k$) to align their internal posterior with the updated social prior, thereby internalizing social norms.

C. Experimental Setup

• Network: A "Caveman Graph" with 14 agents divided into two clusters (0–6 and 7–13) with inter-cluster "hub" agents.
• Conditions:
  • w/ creation: Full model (creation + memorization + communication).
  • w/o creation: Agents only communicate; no novel creations are generated or selected.
• Metrics: Wasserstein distance (point-wise similarity), Gromov–Wasserstein distance (structural similarity), and Representational Similarity Analysis (RSA).

3. Key Contributions

1. Unified Framework: The first computational model to formalize the dialectic of social reality construction within a single Active Inference framework, treating conformity and creativity as complementary variational objectives.
2. Endogenous Group Formation: Demonstrated that informationally cohesive social groups emerge naturally, mirroring the underlying network topology without pre-defined group labels.
3. Circular Mutual Constitution: Showed that social representations and the observation distribution co-evolve. Creative acts project representational structure onto the external world, which in turn reshapes future social representations.
4. Heterogeneous Propagation: Revealed that while social representations diffuse stably within clusters, creative artifacts propagate via selective, dynamic pathways, allowing agents to construct unique "cultural niches."

4. Key Results

A. Emergence of Social Clusters

• Representational Alignment: Over time, agents within the same network cluster developed highly similar social representations, while inter-cluster divergence increased.
• Two-Stage Divergence:
  1. Value Divergence: Agents first diverged in the specific values assigned to observations (Wasserstein distance showed block-diagonal structure early).
  2. Structural Divergence: The relational geometry of representations (Gromov–Wasserstein) diverged later, indicating that local accumulation of novel observations eventually reshaped the global representational space.
• Hub Behavior: Agents bridging the two clusters (hubs) maintained intermediate representational structures, oscillating between the two group norms rather than converging to one.

B. Co-Constitution of Reality

• Structural Similarity (RSA): In the w/ creation condition, the structural similarity between memorized observations and social representations remained high (~0.65). In the w/o creation condition, this similarity decayed monotonically.
• Implication: Creative acts are essential for maintaining the alignment between internal models and the external world. Without active creation, social representations lose their grounding in the observed reality.

C. Propagation Dynamics

• Social Representations: Propagated stably and reciprocally within clusters, creating a "pressure" for internalization that maintains group coherence.
• Creations: Propagated via dynamic, shifting pathways. Creations were often shared between specific pairs (including cross-cluster hubs) rather than diffusing uniformly. This indicates that agents actively construct cultural niches through local interaction dynamics rather than passive diffusion.

5. Significance and Conclusion

This paper provides a rigorous computational explanation for Mead's "I" and "Me" and Berger & Luckmann's dialectic of externalization/objectivation/internalization.

• Theoretical Impact: It resolves the tension between social conformity and individual creativity by framing them as opposing yet complementary forces in free energy minimization. Conformity (model update) provides stability, while creativity (active inference) provides epistemic novelty.
• Mechanism of Social Reality: The study proves that shared social reality is not a static consensus but a dynamic equilibrium maintained by the continuous interplay of internalizing norms and projecting novel observations.
• Future Directions: The authors suggest extending the model to include dynamic network structures (active partner selection) and embodied agents to fully capture the causal chain from intention to physical artifact.

In summary, the model demonstrates that social reality is an emergent property of agents actively minimizing surprise while simultaneously seeking to reduce uncertainty through creative exploration, leading to the spontaneous formation of distinct cultural groups and the continuous evolution of shared norms.