💻 computer science

Emergence of Social Reality of Emotion through a Social Allostasis Model with Dynamic Interpretants

This paper presents a computational model demonstrating that social reality in emotion emerges through the convergence of two agents' interoceptive priors and symbol interpretations, achieved via active inference and dynamic symbol exchange within a social allostasis framework.

Original authors: Kentaro Nomura, Yushi Tsubamoto, Takato Horii

Published 2026-05-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kentaro Nomura, Yushi Tsubamoto, Takato Horii

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine two people, let's call them Alex and Blake, who are trying to figure out how to feel good, but they can't speak the same language yet. They don't have words for "hot," "cold," "hungry," or "tired." Instead, they only have internal signals (like a stomach rumble or a shiver) and a shared goal: to make their internal signals feel comfortable.

This paper describes a computer simulation of how Alex and Blake might invent a shared language for emotions and feelings just by interacting with each other.

Here is the story of how they do it, broken down into simple concepts:

1. The Internal "Thermostat" (Allostasis)

Think of your brain as a smart thermostat for your body. It constantly predicts what your body needs (like energy or temperature) and tries to keep things balanced.

The Problem: Alex's thermostat likes it warm and cozy. Blake's thermostat likes it cool and crisp.
The Action: To feel good, Alex wants to "eat" (to get energy) or "warm up." Blake wants to "cool down."
The Conflict: If they just acted alone, they would never agree on what to do.

2. The "Gestures" (Symbols)

Since they can't speak, they have to invent a way to communicate. They start using symbols (like abstract words or gestures) to say, "I think we should do this action."

At first, their symbols are random. Alex might use a "Blue Dot" to mean "Eat," while Blake uses a "Red Dot" for the same thing.
They play a game called the Naming Game. One person proposes a symbol (the Speaker), and the other person decides whether to accept it or reject it (the Listener).

3. The Magic of "Rejection" (The Learning Mechanism)

This is the most important part of the story. How do they learn to agree?

The Scenario: Alex proposes a symbol meaning "Let's Eat." Blake rejects it because Blake is actually trying to "Cool Down."
The Lesson: Instead of just arguing, Alex learns a lesson. Alex thinks, "Oh, my idea didn't work for Blake. Maybe I should change what I want so that my idea works for both of us."
The Shift: Alex slowly adjusts their internal "thermostat." Instead of wanting it super hot, Alex starts wanting a "medium" temperature. This is a compromise.

Over thousands of rounds of this game, Alex and Blake stop fighting. They both adjust their internal goals until they meet in the middle. They now share the same "ideal state."

4. The Emergence of "Social Reality"

Eventually, something magical happens. The symbols they use stop being random.

A specific symbol (let's say, a "Green Circle") now reliably means "Let's do the action that makes us both feel good."
Because they both agree on what the symbol means, and they both agree on what feeling is "good," they have created a Social Reality.

In this paper, "Social Reality" isn't a physical place; it's a shared agreement. It's the moment when two people stop seeing the world through their own eyes and start seeing it through a shared lens. The emotion (or feeling) isn't just inside Alex's head or Blake's head anymore; it exists in the space between them, built on their mutual agreement.

The Key Takeaways from the Simulation

Emotions are built, not born: The paper suggests that emotions aren't hardwired biological facts (like a fixed "fear" button). Instead, they are like a shared dialect that two people invent together to manage their bodies.
Compromise creates meaning: The meaning of a word (or emotion) changes as the relationship changes. Early in the simulation, a symbol might mean "Eat because I'm hungry." Later, as they get used to each other, that same symbol might shift to mean "Rest because we are both calm."
No telepathy needed: They didn't need to read each other's minds. They just needed to listen to what the other person accepted or rejected, and adjust their own internal goals accordingly.

Summary

Imagine two dancers who start with different rhythms. One wants to dance fast; the other wants to dance slow. They don't have a script. But through trial and error, if one tries a fast move and the other steps back, the first dancer slows down. Eventually, they find a perfect, shared rhythm.

The paper argues that emotions are that shared rhythm. They emerge when we negotiate our internal needs with others, creating a common language of feeling that allows us to live together in a shared reality.

Technical Summary: Emergence of Social Reality of Emotion through a Social Allostasis Model with Dynamic Interpretants

1. Problem Statement

The theory of constructed emotion posits that emotions are not biologically fixed categories but are dynamically constructed through the interplay of bodily regulation (allostasis) and social interaction. A critical gap exists in current computational literature: while some models address the bodily regulation of interoceptive states (e.g., the Embodied Predictive Interoception Coding or EPIC model) and others address the social sharing of concepts via symbol emergence (e.g., the Metropolis–Hastings Naming Game or MHNG), no existing framework integrates both. Specifically, there is a lack of models that demonstrate how the "social reality of emotion"—defined as a community-level consensus on emotion concepts assigned to interoceptive sensations—emerges when agents simultaneously regulate their bodily states and negotiate symbol meanings.

2. Methodology

The authors propose a computational model involving two agents (A and B) that interact to co-construct emotion concepts while regulating their internal bodily states. The system integrates Active Inference for bodily control with Collective Predictive Coding for symbol emergence.

2.1 Agent Architecture

Each agent operates as a Partially Observable Markov Decision Process (POMDP) with discrete states and observations:

Bodily State ( $s^*$ ): Represented by two modalities: energy and body temperature (6 levels each, totaling 36 states).
Observations ( $o^*$ ): Interoceptive signals corresponding to the bodily state.
Actions ( $a^*$ ): Five discrete actions (Cool, Warm, Eat, Play, Sleep) that alter energy and temperature.
Symbols ( $w^*$ ): A shared vocabulary of 15 symbols representing inferred emotion concepts.

2.2 Generative Model Parameters

The agents share and learn specific parameters:

Likelihood ( $A^*$ ): Maps states to observations.
State Transition ( $B^*$ ): Maps actions to state transitions.
Prior Preference ( $C^*$ ): Represents the agent's desired interoceptive state (the "goal" of allostasis).
Symbol Interpretation ( $E$ ): A mapping from symbols to actions (the "meaning" of the symbol).

2.3 Interaction Mechanism

The interaction proceeds in discrete time steps involving three core processes:

Active Inference for State and Symbol Inference: Agents infer their current state and potential symbols to minimize variational free energy. The expected free energy ( $G$ ) is calculated based on the divergence between predicted observations and prior preferences.
Symbol Communication (MHNG): Agents engage in a Metropolis–Hastings Naming Game. One agent acts as a "Speaker" proposing a symbol, and the other as a "Listener." The Listener accepts or rejects the symbol based on a probability derived from its own internal model (specifically, whether the symbol leads to actions that minimize its own expected free energy).
Parameter Updates:
- Likelihood ( $A^*$ ): Updated via Bayesian inference to minimize prediction error.
- Prior Preference ( $C^*$ ): Updated only when a Speaker's symbol is rejected by the Listener. The Speaker shifts its preference toward the observation distribution predicted by the accepted symbol (the Listener's inferred action) and away from its own rejected proposal. This allows agents to adapt their bodily goals to the other without direct access to the other's internal state.
- Symbol Interpretation ( $E$ ): Updated to ensure that the symbol sampled in the MHNG leads to the action that minimizes expected free energy.

3. Key Contributions

Integrated Model: The paper presents the first computational model that unifies the bottom-up process of interoceptive regulation (allostasis) with the top-down process of social symbol emergence.
Mechanism for Social Consensus: It demonstrates a mechanism where "social reality" (aligned prior preferences and shared symbol meanings) emerges solely through the rejection/acceptance dynamics of symbolic communication. Agents adapt their internal goals ( $C^*$ ) to accommodate the other, leading to a compromise.
Dynamic Semantics: The model shows that symbol meanings are not static; they evolve dynamically as the agents' internal states and preferences converge.

4. Experimental Results

The experiment involved two agents starting with divergent prior preferences (Agent A preferred high body temperature; Agent B preferred low body temperature) and no prior knowledge of symbol meanings.

Convergence of Prior Preferences: Over 2.5 million iterations, the Jensen–Shannon (JS) divergence between the agents' prior preference distributions decreased significantly. The agents' preferences converged to an intermediate configuration, indicating a mutual compromise where both agents' goals could be satisfied simultaneously.
Emergence of Symbol Meaning:
- Early Phase: Symbols primarily encoded actions for temperature regulation (Cool, Warm) and energy replenishment (Eat), reflecting the initial conflict in preferences.
- Late Phase: As preferences aligned and the need for temperature regulation decreased, the meaning of certain symbols shifted. Specifically, symbols previously associated with "Eat" shifted to represent "Sleep," reflecting the reduced energy demands of the aligned state.
Validation: The results confirm that agents can align their bodily control goals and co-construct shared emotion concepts without direct access to each other's internal parameters, relying only on the feedback loop of symbol acceptance and rejection.

5. Significance and Claims

The paper claims to provide a computational basis for the "social reality of emotion" as described in the theory of constructed emotion. It demonstrates that:

Social Consensus is Co-Constructive: Emotion concepts are not merely learned but are actively constituted through the negotiation of bodily control goals.
Allostasis is Socially Mediated: The regulation of bodily states is not purely individual; the "body budget" is adjusted based on social interaction (symbol rejection) to achieve a shared reality.
Dynamic Interpretants: The meaning of emotion symbols is fluid, changing in response to the collective state of the interacting agents.

The authors note limitations, specifically that the model is restricted to a discrete environment with a simple configuration. They state that future work will aim to extend the model to continuous environments to test the robustness and scalability of the social reality emergence mechanism.