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

Researcher Aishik Sanyal has published a paper proposing a novel architecture for artificial life agents that demonstrates genuine prosocial behavior through internal coupling rather than external rewards. The work builds on ReCoN-Ipsundrum, an inspectable recurrent controller, by adding an explicit homeostat and a social coupling channel. Crucially, the agent's planning remains strictly self-directed—it only scores its own predicted internal state, with no partner-welfare reward term introduced. This creates a minimal system where helping behavior can be cleanly interpreted.

In experiments across two toy worlds, the architecture showed dramatic results. In the one-step FoodShareToy environment, an exact solver found a sharp behavioral switch from EAT to PASS at a coupling parameter λ* ≈ 0.91. Experimental runs revealed that self-only and partner-observing conditions never produced helping, while affectively coupled conditions always did. In the more complex multi-step SocialCorridorWorld, coupling flipped help rates and partner recovery from 0 to 1, while cutting rescue latency in half from 18 to 9 steps and raising mutual viability from 0.15 to 0.33.

The research employed lesion studies to validate the mechanism: sham lesions preserved helping behavior, but coupling-off and shuffled-partner lesions abolished it completely in both tasks. A parameter sweep revealed a load-dependent feasibility boundary—under low cognitive load, helping appeared for λ ≥ 0.25, but under medium and high loads, no tested parameter value enabled rescue within the time horizon. The paper makes a narrow but significant claim: in this minimal architecture, prosocial behavior emerges specifically when another agent's need is routed into the helper's own self-regulation system.