Empathic Coupling of Homeostatic States for Intrinsic Prosociality

TL;DR

The paper investigates how prosocial behavior can emerge in autonomous agents whose behavior is driven by homeostatic self-regulation. It introduces a homeostatic reinforcement learning framework and contrasts cognitive versus affective empathy in a multi-agent setting, showing that prosocial sharing only arises when agents' internal states are affinely coupled (affective empathy). The experiments span a minimal food-sharing toy and mobile dynamic environments, confirming the necessity of inter-agent homeostatic coupling for sustained helping; cognitive empathy alone does not suffice. The work contributes to understanding intrinsic motivation and provides design principles for socially capable AI, with potential links to neural mirroring mechanisms.

Abstract

When regarding the suffering of others, we often experience personal distress and feel compelled to help. Inspired by living systems, we investigate the emergence of prosocial behavior among autonomous agents that are motivated by homeostatic self-regulation. We perform multi-agent reinforcement learning, treating each agent as a vulnerable homeostat charged with maintaining its own well-being. We introduce an empathy-like mechanism to share homeostatic states between agents: an agent can either \emph{observe} their partner's internal state (cognitive empathy) or the agent's internal state can be \emph{directly coupled} to that of their partner's (affective empathy). In three simple multi-agent environments, we show that prosocial behavior arises only under homeostatic coupling - when the distress of a partner can affect one's own well-being. Our findings specify the type and role of empathy in artificial agents capable of prosocial behavior.

Figures (8)

• Figure 1: A: An experimental setup used in behavioral experiments to test the altruism of monkeys (illustration created based on de1997food). B: A minimal reinforcement learning environment inspired by the behavioral experiment.
• Figure 2: Learning and behavior evaluation in a food sharing environment. A: Learning curves with performance measured by episode duration with both agents alive (n=20, 95% confidence intervals). Only the conditions that implement affective empathy (Affective and Full) result in long episode durations. B: PASS behavior selection rate out of 1,000 steps of the test run. Possessor agents in the Affective condition learn to frequently pass food to the Partner. C: Count of PASS actions when Partner is in the Low energy state. Possessor agents in the Full empathy condition learn to selectively pass food to the Partner when it is most needed. D: LOW state rate of Partner agent, out of 1,000 steps of the test run.
• Figure 3: Overview of the mobile agent environments. A: Linear grid environment. B: 2-D field environment. Detailed explanations are in Appendix \ref{['sec:append_grid']} and \ref{['sec:append_2d']}, respectively.
• Figure 4: Performance in the linear grid mobile environment. A: Learning curves with performance measured by episode duration with both agents alive (n=20). B: Homeostatic drives of agents ($D_{\rm possessor}$ and $D_{\rm partner}$) averaged over 1000 timesteps.
• Figure 5: Performance in the 2-D field mobile environment. A: Learning curves with performance measured by episode duration with both agents alive (n=20). B: An example of helping behavior observed in the Affective condition. The action sequence progresses in order of the numbers in the top right corner of each panel.