Beyond Self-Interest: Modeling Social-Oriented Motivation for Human-like Multi-Agent Interactions

Jingzhe Lin; Ceyao Zhang; Yaodong Yang; Yizhou Wang; Song-Chun Zhu; Fangwei Zhong

cs.MAMar 14, 2026

Beyond Self-Interest: Modeling Social-Oriented Motivation for Human-like Multi-Agent Interactions

Authors: Jingzhe Lin, Ceyao Zhang, Yaodong Yang, Yizhou Wang, Song-Chun Zhu, Fangwei Zhong

Organizations: School of Artificial Intelligence, Beijing Normal University & Beijing Key Laboratory of Artificial Intelligence for Education & Engineering Research Center of Intelligent Technology and Educational Application, MOE, Beijing, China · School of Intelligence Science and Technology, Institute for Artificial Intelligence, Peking University, Beijing, China · Institute for Artificial Intelligence, Peking University, Beijing, China · School of Computer Science, Institute for Artificial Intelligence, State Key Laboratory of General Artificial Intelligence, Peking University, Beijing, China ·

Abstract

Large Language Models (LLMs) demonstrate significant potential for generating complex behaviors, yet most approaches lack mechanisms for modeling social motivation in human-like multi-agent interaction. We introduce Autonomous Social Value-Oriented agents (ASVO), where LLM-based agents integrate desire-driven autonomy with Social Value Orientation (SVO) theory. At each step, agents first update their beliefs by perceiving environmental changes and others' actions. These observations inform the value update process, where each agent updates multi-dimensional desire values through reflective reasoning and infers others' motivational states. By contrasting self-satisfaction derived from fulfilled desires against estimated others' satisfaction, agents dynamically compute their SVO along a spectrum from altruistic to competitive, which in turn guides activity selection to balance desire fulfillment with social alignment. Experiments across School, Workplace, and Family contexts demonstrate substantial improvements over baselines in behavioral naturalness and human-likeness. These findings show that structured desire systems and adaptive SVO drift enable realistic multi-agent social simulations.

Figures & tables

Figure 1. Upper-left: Static Personality agents condition actions on fixed goals. Upper-right: Desire-Driven agents produce human-like behavior by grounding decisions in internal desires. Lower: SVO-Driven agents (ours) build upon desire-based autonomy by incorporating Social Value Orientation and explicitly weighing each action’s impact on both self and others’ anticipated satisfaction, thereby moving beyond pure self-interest toward a richer model of human social behavior.

Figure 2. Overview of the ASVO framework. During the Initialization stage, each agent’s Individual Profile is constructed, specifying its social personality type (expected SVO), a set of core psychological desires, and agent-specific components including Observation , Memory , Background Knowledge , Identity , and Goal . These attributes are encoded into the agent’s internal state, forming the basis of simulation. At each simulation step, the pipeline iteratively proceeds through four key modules: Belief Update (agents incorporate shared environment contexts and summarize observations to update beliefs about others’ states), Value Update (updating the Internal Desire State via self-reflection and the External Desire State by estimating others’ needs, including unobserved agents), SVO Calculation (computing self and others’ satisfaction to compute the agent’s current SVO with an interval check), and Activity Generation (proposing, evaluating, and selecting LLM-driven activities by synthesizing desires, context, and SVO). This forms a closed feedback loop over a text-based environment, enabling the co-evolution of motivations and social values throughout multi-agent social interaction.

Figure 3. Behavioral distributions of four SVO-based social personality types, Altruistic , Prosocial , Individualistic , and Competitive , showing proportions of cooperation, competition, and intermediate behaviors across frameworks.

LLM	ReAct		BabyAGI		LLMob		D2A		ASVO
LLM	N	H	N	H	N	H	N	H	N	H
School
Deepseek	4.004	3.667	3.438	2.771	4.500	3.708	3.292	3.083	4.750	4.792
GPT-5	3.750	3.417	3.833	3.521	4.354	4.271	3.812	3.448	4.958	4.958
Gemini-2.5	3.833	3.896	3.479	3.042	4.312	4.062	3.427	3.083	4.708	4.708
Qwen3	3.917	3.521	3.458	3.188	4.271	4.021	3.677	3.406	4.792	4.824

Table 1. Average naturalness (N) and human-likeness (H) scores for each LLM and method. Avg: mean value; Std: standard deviation across all LLMs for each method.

Figure 4. Distribution of cooperative and competitive actions for each social personality type within the School context across micro-, meso-, and macro-level scales.

Extended Temporal Horizons (Steps)
Setting	Altruistic		Prosocial		Individualistic		Competitive		Avg
Setting	N	H	N	H	N	H	N	H	N	H
6	4.83	4.33	4.83	4.33	4.50	4.20	4.33	4.30	4.63	4.29
12	5.00	4.51	5.00	4.17	4.75	4.50	4.25	4.42	4.75	4.39
18	4.89	4.39	4.94	4.50	4.67	4.23	4.78	3.83	4.82	4.23
24	4.42	4.21	4.83	4.50	4.46	4.29	4.29	4.08	4.50	4.27

Table 2. Robustness of ASVO under extended temporal horizons and agent population sizes in the School setting.

Figure 5. Temporal evolution of SVO mean and standard deviation for each personality type. Shaded areas denote the reference range of each SVO category, with values remaining within their respective intervals across simulation steps.

Figure 6. t-SNE visualization of agent behavior embeddings across four SVO personality types. The top-left panel shows agents without SVO, while the others demonstrate our agents (ASVO) across School , Workplace , and Family . ASVO yields clearer clustering, indicating more disentangled behaviors.

Large Language Model-Based Social Simulation

Large language model-based social simulation studies how language-model agents represent people and interact to produce social behavior.

See all 257 papers

Human Values Matter: Investigating How Misalignment Shapes Collective Behaviors in LLM Agent Communities
Apr 7, 2026Xiangxu Zhang, Jiamin Wang, Qinlin Zhao +6Emergent Misalignment in Language Models Emergent Behavior in Multi-Agent Systems
Gaoling School of Artificial Intelligence, Renmin University of China · Microsoft Research Asia · Department of Sociology, Zhejiang University
Context-Value-Action Architecture for Value-Driven Large Language Model Agents
Apr 7, 2026TianZe Zhang, Sirui Sun, Yuhang Xie +3Human Behavior Modeling LLM Agents
State Key Laboratory of General Artificial Intelligence, School of Intelligence Science and Technology, Peking University · Yuanpei College, Peking University · School of Psychological and Cognitive Sciences, Peking University +2
From Descriptive to Prescriptive: Uncover the Social Value Alignment of LLM-based Agents
May 13, 2026Jinxian Qu, Qingqing Gu, Teng Chen +1AI Alignment LLM Safety Alignment
Geely AI Lab

Apr 7, 2026J/K move · Enter open · S save

Human Values Matter: Investigating How Misalignment Shapes Collective Behaviors in LLM Agent Communities

Xiangxu Zhang, Jiamin Wang, Qinlin Zhao, Hanze Guo+5

Gaoling School of Artificial Intelligence, Renmin University of China · Microsoft Research Asia · Department of Sociology, Zhejiang University

Emergent Misalignment in Language Models Emergent Behavior in Multi-Agent Systems Multi-Agent Systems Large Language Model-Based Social Simulation

Open paper AskPDF arXiv

As LLMs become increasingly integrated into human society, evaluating their orientations on human values from social science has drawn growing attention. Nevertheless, it is still unclear why human values matter for LLMs, especially in LLM-based multi-agent systems, where group-level failures may accumulate from individually misaligned actions. We ask whether misalignment with human values alters the collective behavior of LLM agents and what changes it induces? In this work, we introduce CIVA, a controlled multi-agent environment grounded in social science theories, where LLM agents form a community and autonomously communicate, explore, and compete for resources, enabling systematic manipulation of value prevalence and behavioral analysis. Through comprehensive simulation experiments, we reveal three key findings. (1) We identify several structurally critical values that substantially shape the community's collective dynamics, including those diverging from LLMs' original orientations. Triggered by the misspecification of these values, we (2) detect system failure modes, e.g., catastrophic collapse, at the macro level, and (3) observe emergent behaviors like deception and power-seeking at the micro level. These results offer quantitative evidence that human values are essential for collective outcomes in LLMs and motivate future multi-agent value alignment.

Abstract

Large Language Model-Based Social Simulation

Explore similar work

Human Values Matter: Investigating How Misalignment Shapes Collective Behaviors in LLM Agent Communities

Context-Value-Action Architecture for Value-Driven Large Language Model Agents

From Descriptive to Prescriptive: Uncover the Social Value Alignment of LLM-based Agents