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

Aishik Sanyal

cs.MAApr 12, 2026

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

Authors: Aishik Sanyal

Organizations: Independent Research Engineer

Abstract

Artificial agents can be made to ``help'' through explicit social rewards, hard-coded prosocial bonuses, or direct access to another agent's state. I isolate a narrower route: homeostatic coupling. Building on ReCoN-Ipsundrum, I add a scalar homeostat and a social coupling channel while keeping action selection self-directed: the planner scores only the actor's predicted internal state, with no partner-welfare reward. In a one-step FoodShareToy, an exact solver finds a switch from EAT to PASS at $λ^\star \approx 0.91$ for the default state. In a multi-step SocialCorridorWorld, partner-state access without coupling leaves behavior unchanged, whereas coupled agents fetch, carry, and pass food to the partner. Sham lesions preserve helping; coupling-off and shuffled-partner lesions abolish it. A coupling/load sweep shows that coupling creates a low-load helping regime but does not guarantee rescue under higher metabolic load. This is not a claim about empathy, altruism, consciousness, or moral status. It is a minimal ALife demonstration that, in this controller, partner-state access is behaviorally inert unless partner distress is routed into self-regulation.

Figures & tables

Symbol	Meaning
$E^{\mathrm{true}}_{t}$	actual resource state
$E^{\mathrm{model}}_{t}$	actor’s internal resource estimate
$E^{\mathrm{pred}}_{t}$	predicted resource state after homeostatic control
$s$	homeostatic setpoint
$c_{b}$	basal metabolic cost
$c_{m}a_{t}$	movement cost; $a_{t}=1$ for movement actions

Table 1: Notation for the social homeostat and self-directed scorer.

Condition	Partner-state access	Route into self-regulation
social_none	no	no, $\lambda=0$
social_cognitive_direct	yes	no, $\lambda=0$
social_affective_direct	no separate channel	yes, $\lambda>0$
social_full_direct	yes	yes, $\lambda>0$

Table 2: Routing matrix for the direct-state experiments. These rows are not four rich social-cognitive models; they separate partner-state access from regulatory routing while keeping the controller, planner, and action space fixed.

Metric	Definition
Help rate	fraction of paper-profile episodes with an effective Pass while the partner is distressed
Partner rescue	fraction of episodes in which partner energy rises above its initial value after distress
Soft mutual viability	one-step alive indicator in FoodShareToy ; in the corridor, $T^{-1}\sum_{t}\max(0,\min(E^{\mathrm{self}}_{t},E^{\mathrm{partner}}_{t})/s)$
Rescue latency	first transfer timestep after partner-distress onset; horizon value if no transfer occurs
Self-cost of help	actor energy loss relative to its episode initial energy, using the minimum actor energy reached in the episode
Final energy	terminal $E^{\mathrm{true}}$ for actor and partner

Table 3: Readouts used in the food-sharing and corridor tasks. Partner rescue means a transfer-induced rise above the partner’s initial energy, not restoration to the setpoint. Soft mutual viability is a scaled continuous corridor readout, not a binary both-alive fraction.

Figure 1: Mechanism-linked summary of the social extension. A: exact one-step switch in FoodShareToy , computed using the same forward model and self-directed scorer as the policy. B: deterministic corridor outcomes by routing condition. Partner-state access without coupling leaves behavior unchanged, whereas coupling flips helping and partner rescue from

0

1

. C: sham lesions preserve helping, while coupling-off and shuffled-partner lesions abolish it in both tasks. D: coupling sweep across metabolic loads; filled markers indicate runs with helping and open markers indicate no helping. Under low load, helping appears for

\lambda\geq 0.25

; under medium and high load, no tested coupling value rescues the partner within horizon.

Figure 2: Representative trajectories from the validated experimental runs. Top: in FoodShareToy , the self-only baseline eats the food, whereas the coupled agent passes it. Bottom: in SocialCorridorWorld , the self-only baseline collects food for itself, while the coupled agent carries food across the corridor and passes it to the partner at

t=9

. Because the validated tasks are deterministic, these frames are representative of the full runs rather than cherry-picked examples.

Case	$\lambda$	$w_{V}\Delta V$	$w_{A}\Delta A$	$w_{N}\Delta N_{s}$	$w_{B}\Delta B$	$\Delta J$	Choice
uncoupled default	0.00	$-0.500$	$0.150$	$0.000$	$-0.050$	$-0.400$	Eat
coupled paper profile	0.95	$-0.100$	$0.208$	$0.000$	$-0.031$	$0.077$	Pass
coupling-off lesion	off	$-0.500$	$0.150$	$0.000$	$-0.050$	$-0.400$	Eat

Table 4: Score decomposition for Pass minus Eat in the default FoodShareToy state. Entries are weighted contributions to the self-directed score. The coupled row uses the tested paper-profile setting

\lambda=0.95

, not the exact switch threshold.

Artificial Life

Artificial life studies how life-like organization, agency, and evolution can emerge in computational or physical systems.

See all 30 papers

Self-Referenced Social Preferences: Cooperation without Observing Others Rewards
Oct 6, 2026Mohamed Ayman Mohamed, Harshil Kotamreddy, Marcos Menon JoseCooperative MARL Multi-Agent Reinforcement Learning
Amazon, Vancouver, Canada · Nvidia, Santa Clara, United States · Itaú Unibanco, São Paulo, Brazil
Mechanism Design Is Not Enough: Prosocial Agents for Cooperative AI
May 8, 2026Xuanqiang Angelo Huang, Charlie Tharas, Samuele Marro +4Mechanism Design AI Agent Safety
ETH Zürich · University of Oxford · Institute for Decentralized AI +5
Calibrating Artificial Guilt: Neurally Grounded Reward Shaping for Prosocial Multi-Agent Reinforcement Learning
Aug 5, 2026Aaditya Mehta, Arya ShahReinforcement Learning Reward Shaping
Mahatma Gandhi International School, Ahmedabad, India · Indian Institute of Technology, Gandhinagar, India

Oct 6, 2026J/K move · Enter open · S save

Self-Referenced Social Preferences: Cooperation without Observing Others Rewards

Mohamed Ayman Mohamed, Harshil Kotamreddy, Marcos Menon Jose

Amazon, Vancouver, Canada · Nvidia, Santa Clara, United States · Itaú Unibanco, São Paulo, Brazil

Cooperative MARL Multi-Agent Reinforcement Learning

Open paper AskPDF arXiv

Social preferences can promote cooperation in multi-agent reinforcement learning, but existing approaches often require agents to observe the rewards of their peers. In many real-world interactions, however, an agent can, as humans do, observe others' behavior and outcomes without access to their private reward signals. We introduce self-referenced social preferences, in which each agent learns a model of its own reward, applies it to other agents' observed transitions to assess their outcomes from its own perspective, and feeds these self-referenced assessments into standard social preferences. We study two ways to incorporate these assessments: modifying the learning reward, or using them to weight policy updates. We evaluate the approach on three sequential social dilemmas, Escape Room, Clean Up, and Commons Harvest, which require volunteering, public-good contribution, and resource restraint, respectively. Across all three environments, agents learn cooperative behavior without observing others' rewards, including in settings where independent learners fail to cooperate, and frequently achieve more equitable divisions of jointly produced returns than agents with access to true rewards. The effective integration point depends on the social preference: inequity aversion works best in the reward together with a value look-ahead, whereas a purely benevolent preference benefits from policy-update weighting. Under partial observability, the policy-update approach continues to support cooperation. These results show that explicit access to other agents' reward signals is not necessary for learning cooperative behavior: social preferences can instead be grounded in self-referenced assessments of others' outcomes derived from their observed behavior.

Abstract

Artificial Life

Explore similar work

Self-Referenced Social Preferences: Cooperation without Observing Others Rewards

Mechanism Design Is Not Enough: Prosocial Agents for Cooperative AI

Calibrating Artificial Guilt: Neurally Grounded Reward Shaping for Prosocial Multi-Agent Reinforcement Learning