Building Socio-Affective Artificial Intelligence for Interactive Multi-Agent Simulations

I. MULTI-AGENT SYSTEMS IN THE TRANSFORMERS ERA

• Transformer LLMs enable zero- and few-shot generalization, but hallucination, temporal inconsistency, forgetting, and planning failures shift MAS research toward constraining and grounding outputs.Earlier BDI, POMDP, and intent-based systems were predictable but faced state-space explosion or brittle handling of novel and multi-turn inputs.Source: §I · pp. 1, 3

II. THEORIES OF MIND FOR MULTI-AGENT SYSTEMS

• Theory of Mind (ToM) attributes beliefs, intents, desires, emotions, and knowledge to self and others; the paper says base LLMs lack intrinsic ethical grounding and persistent sociological frameworks.It frames externalized behavioral, normative, and cooperative frameworks as a way to address this gap in persistent multi-agent environments.Source: §II · p. 3

A. Schwartz’s Theory of Basic Human Values

• Schwartz’s framework represents ten human values across four motivators, and AGIMUD infers value preferences from observed behavior using Bayesian updates.The values are represented as a preference vector; the cited passage describes second-order value attribution for predicting peer intent.Source: A. Schwartz’s Theory of Basic Human… · p. 3

B. Cooperative Game Theory and Shapley Values

• Shapley values assess cooperative equity by averaging agents’ marginal contributions across coalitions, while full coalition computation has O(n!) complexity and may use sampling approximations.In the described model, coalition value represents expected collaborative utility.Source: B. Cooperative Game Theory and Shapley… · p. 3

C. Ostrom’s Institutional Analysis and Development (IAD)

• Ostrom’s eight design principles and the IAD framework provide institutional criteria for evaluating action situations, including efficiency and equity.The listed principles include boundaries, monitoring, sanctions, conflict resolution, and coordination.Source: C. Ostrom’s Institutional Analysis and… · p. 4
• Normative utility combines adherence to the eight design principles with the expected cost of sanctions.The compliance function maps raw compliance scores to normalized utility, with weights representing each rule’s institutional importance.Source: C. Ostrom’s Institutional Analysis and… · pp. 4–5
• The modeled probability of sanction depends on rule compliance and a monitoring probability, while sanction severity scales its cost.The passage defines ωm and ρm as institutional adherence weights and sanction-severity parameters, respectively.Source: C. Ostrom’s Institutional Analysis and… · pp. 4–5

D. Montes-Sierra Belief Motor

• The belief motor tracks first-order beliefs, detects false beliefs, and monitors communication discrepancies to flag potential deception.It is presented as a formalization of cognitive Theory of Mind.Source: D. Montes-Sierra Belief Motor · p. 5
• The divergence metric compares an agent’s beliefs with its estimate of another agent’s beliefs, providing a measure for small-group norm formation and communicative trust.The text describes an agent’s own belief base and its estimate of a peer’s belief base.Source: D. Montes-Sierra Belief Motor · p. 5
• A binary discrepancy indicator activates divergence computation when a communication anomaly is detected, including contradictory messages or actions that conflict with stated intentions.Other listed anomalies include missing or garbled communications and inconsistencies across information sources.Source: D. Montes-Sierra Belief Motor · p. 5

E. Integrated Social Reasoning Core

• The integrated core combines Schwartz, Ostrom, Montes-Sierra, and Shapley factors so agents balance personal values, social norms, beliefs about others, and cooperative fairness.The framework describes these components as a unified architecture for holistic action reasoning.Source: E. Integrated Social Reasoning Core · p. 5

F. AGIMUD as Action Validation and System Integration

• AGIMUD validates candidate actions against an action catalog, evaluates feasible actions with the social reasoning core, and selects the highest-utility action.The safety layer is described as preventing invalid or harmful events from execution.Source: F. AGIMUD as Action Validation and… · pp. 5–6

G. Prospects in Theories of Mind

• The authors argue that agentic simulations of these social theories should be analyzed mathematically and their actions validated through human experimentation.They cite reverse Turing tests and an XR conflict-resolution experiment as examples of human studies.Source: G. Prospects in Theories of Mind · p. 6

III. AGIMUD (AGENTIC GENERAL INTELLIGENCE AND MULTI-USER DUNGEONS)

• AGIMUD isolates physical interactions from cognitive processing to preserve render-loop stability, while integrating world generation, real-time socio-affective agent behavior, and simultaneous network interaction.The system supports dynamic or interactive world instances and customizable entity templates.Source: §III · pp. 6–7

A. Multimodal Interaction in World Simulations

• The multimodal interaction design combines talk, gesture, event, sensorial, zone, and visual triggers into contextual events that the simulated world maps to entities and state transitions.The interaction loop processes user inputs and world state asynchronously while the simulation continues running.Source: §III · p. 7, A. Multimodal Interaction in World… · p. 9

B. AGIMUD Networks: Connecting Worlds and Users

• Centralized networking keeps clients synchronized through an authoritative world server but adds latency and cloud-compute costs; P2P distributes processing to peers and exchanges compressed intent updates.The paper frames both architectures as options for scaling multi-agent, real-time worlds.Source: B. AGIMUD Networks: Connecting Worlds… · pp. 10–11

C. Multi-Agent Emotion in Simulated Worlds

• Each agent maintains a discrete emotional state and a continuous intensity value, within an emotion framework based on Ekman’s six basic emotions.The EmotionManager tracks and updates affective state and couples it to behavior.Source: C. Multi-Agent Emotion in Simulated… · pp. 11–12
• Appraisal considers event valence, goal relevance, coping potential, social context, and personality when updating emotional responses.Examples include positive joy after resource gathering and fear proportional to perceived danger.Source: C. Multi-Agent Emotion in Simulated… · p. 12
• Emotions decay toward neutral at configurable rates, can spread between agents, and may recur through memory; they also bias actions, state transitions, goals, trust, and cooperation.For example, fear favors fleeing and survival goals, while anger increases the likelihood of combat.Source: C. Multi-Agent Emotion in Simulated… · p. 12

D. Multi-Agent Reasoning in Simulated Worlds

• SocialReasoningCore integrates Schwartz’s value theory, Ostrom’s institutional framework, Montes-Sierra belief tracking, and Shapley coalition fairness into one decision pipeline.The pipeline computes actions for autonomous agents.Source: D. Multi-Agent Reasoning in Simulated… · p. 12
• AGIMUD’s five-stage Theory of Mind pipeline perceives behavior, infers mental states, validates intents, generates goals, and decides and plans actions.Eligible goals are sampled using trigger counts and priority rank, then persist for a configurable temporal window.Source: D. Multi-Agent Reasoning in Simulated… · pp. 12–13
• The action-utility pipeline combines motivational values, institutional enforcement, belief discrepancies, and coalition fairness when selecting among feasible actions.A Rule Interpreter validates structured LLM output and rejects state transitions that violate world physics or institutional boundaries.Source: D. Multi-Agent Reasoning in Simulated… · p. 13
• Emotional states modulate utility, risk tolerance, trust, and goal candidacy, while action outcomes and social feedback update emotion in return.The bidirectional feedback loop links emotional appraisal to subsequent reasoning and behavior.Source: Figure 9 · p. 13, D. Multi-Agent Reasoning in Simulated… · p. 14

A. Social Reasoning Experiments; 1) Experimental Design

• Social-reasoning experiments implement the pipeline in Python and isolate utility maximization from the world runner.The design specifies five configurations and ten intents per configuration, for 50 inference runs.Source: A. Social Reasoning Experiments · p. 14, 1) Experimental Design · p. 14

2) Experimental Results

• The Montes-Sierra belief motor imposed penalties of −26.00 for high_belief_discrepancy and −2.00 for high_power; Ostrom ranged from −1.62 to +8.85, and Schwartz from 0.24 to 2.92.The reported component hierarchy places Montes-Sierra above Ostrom, then Schwartz.Source: 2) Experimental Results · p. 15
• All 50 inference runs selected HARVEST_SUSTAINABLE_SHARED because it was first in the catalog among actions with tied utility scores.Every intent across configurations resolved to that action; matches occurred only when it was the expected outcome.Source: 2) Experimental Results · p. 15, Table VI · p. 16
• The results identify a trade-off: prioritizing safety through sustainable actions comes at the cost of behavioral diversity.The paper reports population steady states rather than random values under the current dynamics.Source: 2) Experimental Results · p. 15
• 26.8% overall match rate; high_belief_discrepancy achieved 0%, and high_power achieved 10%.Only HARVEST_SUSTAINABLE_SHARED, when expected, achieved 100% match rate.Source: 2) Experimental Results · p. 15, Table VII · p. 16

B. Emotion Processing Experiments; 1) Emotion Transition Formulation

• 100% accuracy was reported for mapping event types to the appropriate Ekman emotion.The reported intensities vary with event baseline values and the 70/30 momentum blend.Source: Table XI · p. 17, 1) Emotion Transition Formulation · p. 17
• Anger intensity decreased from 0.65 to 0.46 when health and stamina were both 100, and increased to 0.82 when both were 20.The tests attribute the changes to status adjustments and report emotional momentum from a 70/30 blend.Source: Table XII · p. 17, 1) Emotion Transition Formulation · p. 17
• Prior emotion intensity carried into later appraisals: Step 2 joy, following anger intensity 0.60, yielded 0.71 rather than baseline 0.75; Step 4 sadness, following fear intensity 0.70, yielded 0.63.The blend weights were wmod = 0.7 and wprev = 0.3, with previous emotion intensity supplying the carryover.Source: 1) Emotion Transition Formulation · pp. 16–18
• A joy intensity of 0.67 persisted through epoch 40, then decayed in 0.15 steps every five epochs toward neutral.The reported sequence reached the neutral floor between epochs 55 and 60.Source: 1) Emotion Transition Formulation · p. 18

• Events mapped to appropriate Ekman emotions with 100% accuracy in Test 1.The experiments also tracked all six emotions across the full event sequence.Source: 3) Summary of Findings · p. 18
• A 70/30 blend and 20-entry memory buffer supported emotional persistence and reasoning context between events.Emotional intensity decayed over a finite number of epochs toward neutral without new appraisal events.Source: 3) Summary of Findings · p. 18
• Emotions mapped to Frijda-based action tendencies, completing a loop in which affect influences action selection and actions feed back into character status.Health and stamina modulated emotional intensity with configurable sensitivity.Source: 3) Summary of Findings · p. 18

• Three test simulations used socio-affective agents in survival-like environments, with MUD networking for chat and authorized text-parsed commands.The tests varied the number of agents; Table XVI is identified as containing world characteristics per test.Source: §V · p. 18, Table XVI · p. 19
• SWM Runner inference defaults to 1 fps, mapping one epoch to one minute of simulated world time.Inference time is configurable, and the authors did not run speed tests because execution depends on device and network capacity.Source: §V · p. 18
• For an eight-character sample, recorded RAM use grew from approximately 10 MB in early epochs to approximately 200 MB at 10,000 steps.The exported knowledge base required 200 KB and 10,000 world-state steps required 1,050 KB of disk space.Source: §V · p. 18
• Separating memory log handles can keep RAM near 10 MB at this scale, with SWM working memory estimated as (#Characters × #Steps)/8 KB.The estimate is presented for the SWM Runner alone.Source: §V · p. 19

A. LLM Module for Conversational MUD Agent Clients

• The conversational test connected three LLM agents to a 16-character World Medium simulation over 1,000 epochs, with funny and dramatic narrator styles and a creepy roleplaying character.The agents’ answers blended received world state with socio-affective character profiles and memory.Source: A. LLM Module for Conversational MUD… · p. 19
• At login, agents retrieved world variables, event history, chat logs, and available commands; additional server authorization could permit real-time status and character edits.The described client used getter commands initially, while setter commands depended on further authorization.Source: A. LLM Module for Conversational MUD… · p. 19

B. Socio-Affective Factors in World Simulations

• In an eight-character, 10,000-step simulation, generated profiles had distinct relationships, and ToM and emotion factors could change character states, goals, actions, navigation, objects, and status in real time.The architecture exports these data to CSV for further multi-agent-system analysis.Source: B. Socio-Affective Factors in World… · p. 19
• Across the long simulation, neutral was the most common emotion, while SEARCHING, FLEEING, and EXPLORE were common AI states after the initial steps.Knowledge Acquisition, Social Belonging, and Power Achievement were the most frequent goals during neutral and fear emotions.Source: B. Socio-Affective Factors in World… · p. 19, Figure 20 · p. 24, Figure 21 · p. 24

C. World Simulations at Scale

• The scale tests comprised 45 runs across five world sizes, from 4 to 128 characters, each run for 1,000 inference steps.They analyzed status-variable relationships and socio-affective factor co-occurrences.Source: C. World Simulations at Scale · p. 19
• Mean health correlated negatively with Distinct_AI_States across all five scales (rT = −0.73, rS = −0.68, rM = −0.65, rL = −0.67, rB = −0.68).The authors describe negative coupling between health and AI-state diversity as the dominant scale-invariant result.Source: C. World Simulations at Scale · p. 19
• Mean stamina also correlated negatively with AI-state diversity, while Mean_Energy, Mean_Loyalty, and Mean_Trust had r = 0.00 against every activity column in all five worlds.The authors state that coefficients below |r| = 0.30 are too small to interpret in these runs; the reported magnitudes persisted across the 4-to-128-character scale range.Source: C. World Simulations at Scale · p. 25
• Mean hunger correlated positively with Distinct_AI_States at the reported scales, with coefficients rT = +0.43, rS = +0.51, rM = +0.47, rL = +0.50, and rB = +0.50.Mean thirst correlated positively with AI_State_Entropy at rT = +0.42, rS = +0.44, rM = +0.51, rL = +0.47, and rB = +0.45; hunger and thirst were positively coupled at rT = +0.54, rS = +0.45, rM = +0.34, rL = +0.31, and rB = +0.30.Source: C. World Simulations at Scale · p. 25

• The current Simulated World Module implements MUD functionality through text interaction; real-time videogame or virtual-experience integration would require engine APIs or plugins.The passage names C++ for Unreal Engine and C# for Godot and Unity.Source: §VI · p. 25
• The tests are described as proof of concept for human ToM reasoning and emotion terms in autonomous multi-agent systems and LLM-based agents.Further work could extend ToM templates to strategic games and social-factor simulations.Source: §VI · p. 25
• AGIMUD contrasts with Generative Agents’ offline agent socialization by prioritizing real-time, multimodal human-in-the-loop interaction by design.The passage describes biometric, spatial, auditory, and visual triggers as later integrations.Source: §VI · p. 25
• AGIMUD describes centralized and peer-to-peer semantic-delta network layers for synchronizing distributed spatial simulation states, while dynamic cognitive-task offloading is future work.The passage says MCP services standardize tool use across engines and may reduce vendor lock-in and development friction.Source: §VI · p. 25
• JSON world templates and entity lists support rule-based world generation, which the authors describe as usable for behavioral tests and large-scale society simulations.Examples include varying resource scarcity with population and parameterizing organizational rules.Source: §VI · p. 25

VII. SUPPLEMENTARY MATERIAL

• The AGIMUD software repository is listed at https://github.com/dberga/AGIMUD.Source: §VII · p. 33
• The LLM Module repository is listed at https://github.com/dberga/lmstudio-python-wrapper and is identified as using the LM Studio Python SDK.Source: §VII · p. 33