Prosociality by Coupling, Not Mere Observation: Homeostatic Sharing in an Inspectable Recurrent Artificial Life Agent

1. Problem Statement

In synthetic artificial life (ALife) systems, "prosociality" (helping behavior) is often trivialized or difficult to interpret because agents may help due to:

• Explicit social reward terms in the objective function.
• Hard-coded bonuses for helping.
• Direct access to another agent's internal state without a mechanistic link to the helper's own regulation.

The core question addressed is: Does merely observing another agent's distress cause prosocial behavior, or is it necessary for that distress to be dynamically coupled into the agent's own homeostatic self-regulation? The author aims to decompose this distinction in a minimal, hand-specified, and fully inspectable recurrent controller, avoiding the "black box" nature of learned multi-agent systems.

2. Methodology

Base Architecture: ReCoN-Ipsundrum

The study extends a recurrent controller (ReCoN) augmented with an "ipsundrum" persistence loop and an affect proxy. The key innovation is the addition of an explicit scalar homeostat and a social coupling channel.

• Homeostatic State ($E$): Each agent maintains a true energy state ($E^{true}$) and a predicted model state ($E^{model}$).
• Distress Calculation:
  • $d_{self}$: Distress based on the gap between the setpoint ($s$) and the agent's own predicted energy.
  • $d_{other}$: Distress based on the gap between the setpoint and the partner's estimated energy ($\hat{E}_{other}$).
  • Coupled Distress ($d_{cpl}$): $d_{cpl} = d_{self} + \lambda d_{other}$, where $\lambda$ is the affective coupling strength.
• Prediction & Regulation: The agent predicts its future state based on $d_{cpl}$. The prediction error ($PE$) updates the internal model.
• Planning Constraint (Crucial): The planner is strictly self-directed. It evaluates actions based only on the agent's own predicted internal state (valence, arousal, salience, and body-budget error).
  • No external reward: There is no "help partner" term or separate social welfare bonus in the scoring function.
  • Mechanism of Help: Helping can only occur if the partner's distress ($d_{other}$) is routed into the agent's own coupled distress ($d_{cpl}$), thereby altering the agent's self-regulation score.

Experimental Design

The study compares four matched conditions in two toy worlds:

1. None: No partner access.
2. Cognitive Direct: Direct access to partner state, but $\lambda = 0$ (no coupling).
3. Affective Direct: Partner state routed to homeostat with $\lambda > 0$.
4. Full Direct: Direct access + coupling ($\lambda > 0$).

Environments:

• FoodShareToy: A one-step decision task. An agent with food chooses to EAT or PASS to a passive partner with low energy. An exact solver is used to find the theoretical switching point.
• SocialCorridorWorld: A multi-step sequential task (18 steps). The agent must navigate a corridor, fetch food, avoid hazards, and pass it to a partner.

Interventions:

• Lesions: "Sham" (control), "Coupling-off" ($\lambda=0$), and "Shuffled-partner" (routing incorrect partner data).
• Coupling Sweep: Testing $\lambda \in \{0, 0.25, 0.5, 0.75, 1.0\}$ under low, medium, and high metabolic loads.

3. Key Contributions

1. Mechanistic Decomposition: A minimal, non-learning ALife model that isolates the difference between observation (cognitive access) and coupling (regulatory integration).
2. Exact Solvability: The use of a one-step toy world allows for an exact mathematical derivation of the threshold where behavior switches, validating the empirical results.
3. Causal Lesioning: Demonstrates that removing the coupling channel abolishes helping without changing the action space or reward structure, proving the mechanism is architectural rather than reward-based.
4. Ecological Boundary: Identifies that prosociality is not a monotonic function of coupling strength but depends on the interplay between coupling ($\lambda$) and metabolic load.

4. Results

The Observation vs. Coupling Dissociation

• FoodShareToy:
  • The exact solver identifies a sharp threshold at $\lambda^* \approx 0.91$. Below this, the agent EATs; above it, the agent PASSES.
  • Empirical Runs:
    • None & Cognitive Direct ($\lambda=0$): Help rate = 0. The agent eats the food. Mere observation of the partner's need does not alter the self-directed score.
    • Affective & Full Direct ($\lambda>0$): Help rate = 1. The agent passes the food. The partner's distress is routed into the agent's own homeostatic error, making "PASS" the optimal self-regulation strategy.
• SocialCorridorWorld:
  • None & Cognitive Direct: Help rate = 0. The agent collects food for itself and ignores the partner.
  • Affective & Full Direct: Help rate = 1. The agent successfully executes a fetch-carry-pass sequence.
  • Metrics: Coupling flips partner recovery from 0 to 1, reduces rescue latency from 18 steps (horizon limit) to 9 steps, and increases mutual viability from 0.15 to 0.33.

Lesion Analysis

• Sham Lesions: Helping behavior is preserved.
• Coupling-off & Shuffled-Partner: Helping is abolished in both tasks. The agent reverts to the self-only baseline. This confirms that the behavior relies specifically on the integrity of the coupling channel, not just the presence of partner data.

Load-Dependent Feasibility Boundary

• Low Metabolic Load: Helping emerges at $\lambda \ge 0.25$.
• Medium/High Metabolic Load: No tested value of $\lambda$ results in helping. The metabolic cost of helping exceeds the agent's capacity within the time horizon, regardless of coupling strength.
• Conclusion: Coupling reorganizes the policy toward helping, but success is conditional on ecological parameters (energy availability and time).

5. Significance and Implications

• Philosophical Distinction: The paper provides a rigorous, executable distinction between knowing another is in need (cognitive access) and being affected by that need (regulatory coupling). In this architecture, only the latter produces prosociality.
• Methodological Advance: It demonstrates that tunable, minimal ALife models can serve as "experimental philosophy" tools. By varying parameters ($\lambda$, load, lesions), researchers can test which social descriptions survive mechanistic decomposition.
• Refutation of "Social Reward" Myths: The study proves that prosocial behavior can emerge without explicit social reward terms. It arises purely from the agent's attempt to minimize its own homeostatic error when that error is coupled to another's state.
• Limitations & Future Work: The model is deterministic, uses passive partners, and lacks learning or complex social inference. However, these constraints are intentional to ensure the causal story remains legible. Future work could explore developmental sociality and active partners.

Final Takeaway: In this minimal recurrent architecture, prosociality is not a result of observation, but of coupling. Helping occurs only when another agent's distress is routed into the helper's own self-regulation loop, effectively making the partner's survival a component of the agent's own homeostatic imperative.