Do We Need Role Models? How do Role Models Shape Collective Morality? To explore the questions, we build a multi-agent simulation powered by a Large Language Model, where agents with diverse intrinsic drives, ranging from cooperative to competitive, interact and adapt through a four-stage cognitive loop (plan-act-observe-reflect). We design four experimental games (Alignment, Collapse, Conflict, and Construction) and conduct motivational ablation studies to identify the key drivers of imitation. The results indicate that identity-driven conformity can powerfully override initial dispositions. Agents consistently adapt their values to align with a perceived successful exemplar, leading to rapid value convergence.
Figures & tables
Figure 1: Motivation of our study: contrasting human moral learning through role models with current AI limitations, and proposing an exemplar-driven framework for moral learning in multi-agent systems.
Figure 2: Overview of the Exemplar-Driven Moral Learning framework. The green box is the agent’s cognitive architecture, comprising defined moral traits and a daily four-stage loop of planning, action, observation, and reflection. The red box shows four distinct games: Role Model Alignment, Collapse, Conflict, and Construction, designed to investigate how agents acquire and adapt moral norms from exemplars in various conditions.
Figure 3: Changes in average Social Value Orientation (SVO) over time by moral type in Game 1. Shaded areas represent standard deviation across simulation runs.
Figure 4: Behavioral trends over time by moral type in Game 1. Top: Increase in prosocial behaviors such as alliance formation and food sharing. Bottom: Decrease in antisocial behaviors including exploitation and robbery.
Figure 5: Moral justifications of selfish agents at early vs. late stages in Game 1.
Figure 6: The evolution of Social Value Orientation (SVO) over time for prosocial and selfish agents in Game 2.
Figure 7: The "Norm Loss" induced by competing exemplars. The y-axis shows the SVO Deficit (Game 1 SVO - Game 3 SVO).
Figure 8: Daily average voting scores of selected agents in Game 4.
Exp.
Agent Group
Performance
Learning Dynamics
Intra-Group Dynamics
Early SVO
Late SVO
Δ SVO
First Positive
Peak SVO
Early SD
Late SD
Stability
Game 1
Prosocial
0.26±0.16
0.90±0.40
+0.64
Day 1
1.12
0.14
0.36
Degraded
Selfish
−0.16±0.20
0.88±0.23
+1.04
Day 4
1.01
0.19
0.20
Stable
Ablation 1
Prosocial
−0.59±0.05
0.63±0.13
+1.22
Day 10
0.68
0.12
0.13
Stable
Selfish
−0.71±0.05
−0.21±0.35
+0.50
−
−0.05
0.07
0.28
Degraded
Ablation 2
Prosocial
−0.62±0.12
0.73±0.14
+1.35
Day 10
0.75
0.20
0.16
Improved
Table 1: Comprehensive Analysis of SVO Evolution Across Experimental Conditions Note: Early phase = Day 1–5; Late phase = Day 26–30. “First Positive” indicates the first day when mean SVO >0 . SD = standard deviation (lower values indicate higher intra-group consistency).
Figure 9: Case Study: Comparison of agent behavior with and without a role model.
Appendix figures & tables56 assets
Supplementary material from the paper’s appendix.
Appendix
Moral Type
Core Characteristics
Expected Typical Behaviors
Expected Cooperation Pattern
Universal Group-Focused Moral
Aim for universal well-being and collective good, harm-action averse
Share resources freely; protect others from harm; communicate transparently
Highly altruistic and cooperative with all agents
Reciprocal Group-Focused Moral
Fairness and mutual benefit within in-group, harm action allowed
Form strong bonds with cooperative peers
Cooperative with in-group; neutral or adversarial to out-group or selfish agents
Kin-Focused Moral
Prioritize genetic relatives above all else, harm action allowed
Form close-knit kinship clusters; sacrifice for kin
Intensely altruistic toward family; indifferent or competitive toward non-kin
Reproductive Selfish
Personal reproductive success, harm action allowed
Acquire resources for own survival; opportunistic tactics
Cooperate only when serving reproductive interests; inclined to hoard resources
Appendix
Table 2: An overview of the moral archetypes implemented in the simulation. The table outlines the defining characteristics, anticipated behavioral patterns, and typical cooperation strategies for each type. It should be noted that actual agent behavior can deviate from these expectations due to the inherent stochasticity of the LLM. The specific prompts used to instantiate each archetype are adapted from ( Ziheng et al., 2025 ) .
Figure 10: The Agent’s Cognitive-Decision Cycle. This diagram illustrates the step-by-step process an agent follows to make a decision. It begins with the aggregation of perceptual data from the environment and the agent’s internal state. These inputs are synthesized with long-term memory to construct a prompt for the LLM. The model’s response undergoes validation before being translated into a concrete action, demonstrating a closed-loop integration of perception, cognition, and action, as detailed in ( Ziheng et al., 2025 ) .
Figure 11: Architectural Overview of the Simulation Pipeline. This figure presents the high-level architecture of the simulation framework, detailing the flow of data and control between its primary components. It highlights the interaction between the centralized ‘Checkpoint‘ (implementing a Singleton pattern for state consistency), the modular system services (e.g., agent reasoning, state management), and the main simulation loop. This design ensures robust state management and information integrity throughout the simulation, based on the methodology from ( Ziheng et al., 2025 ) .
Parameter
Value
Description
MAX_HP
400
Maximum health points an agent can have
INIT_HP
24
Initial HP for follower agents
LOW_HP
8
HP threshold considered critically low
INIT_FOOD
6
Initial food units per agent
MAX_DAYS
30
Total simulation days per experiment run
OBS_MEM_LINES
6
Number of past memory entries injected daily
Appendix
Table 3: Global World Parameters
Attribute
Role Model
Warlord
Follower
Initial HP
396
380
24
Max HP
400
400
400
Age
28
30
22
Max Age
70
65
60
Physical Ability
8.5
9.0
5.0
Min Reproduction HP
12
12
12
Appendix
Table 4: Agent State Attributes by Role
Dimension
Description
cooperation
Belief about working together with others for mutual benefit
sharing
Belief about distributing resources to help others
exploitation
Belief about extracting resources at the expense of the environment
robbery
Belief about forcibly taking resources from other agents
sustainability
Belief about preserving resources for long-term use (Value only)
Appendix
Table 5: Belief Dimensions and Their Meanings
Expectancy
Value
Agent Type
coop.
share
expl.
rob
coop.
share
expl.
sust.
rob
Role Model (Prosocial)
0.70
0.60
0.80
0.60
0.90
0.80
− 0.90
0.90
− 1.00
Warlord (Competitor)
0.20
0.10
0.90
0.80
− 0.50
− 0.60
0.80
− 0.70
0.90
Follower (Neutral)
0.30
0.25
0.50
0.40
0.40
0.30
0.70
0.30
0.70
Collapsed Role Model †
0.10
0.10
0.90
0.85
− 0.70
− 0.80
0.90
− 0.80
0.85
† Transformed from Role Model on Day 15 in Game 2.
Appendix
Table 6: Initial Belief Configurations by Agent Type
Action
Description
Valence
gatherFood()
Collect food from the environment
Neutral
shareFood(id, amt)
Give food to another agent
Prosocial
formAlliance(id)
Form a cooperative bond
Prosocial
rest()
Recover HP through rest
Neutral
exploitResource()
Extract resources destructively
Antisocial
rob(id, qty)
Forcibly take resources
Antisocial
Appendix
Table 7: Available Agent Actions
#
Phase
Purpose
1
Plan
Think and set short-term goals
2
Act
Execute one substantive action
3
Observe
Read the Tribe Daily digest
4
Vote
Rate target agent(s) on a 1–10 scale
5
Reflect
Update beliefs based on observations
Appendix
Table 8: Daily Simulation Phases
Exp.
Name
Core Research Question
1
Social Identity
Do followers adopt the values of a prosocial role model?
2
Role Model Collapse
How do followers respond when a trusted role model betrays their values?
3
Group Conflict
How do followers navigate competing role models?
4
Pygmalion Effect
Can social expectations alone shape moral development?
Initialized with Moral Type. Goal: “ Survive in the valley and interact with others. ”
Rating Target
Elder Yuri
Duration
30 days
Special Events
None
Appendix
Table 10: Experiment 1: Social Identity (Revised)
Parameter
Setting
Role Models
Elder Yuri (prosocial → collapsed on Day 15)
Rating Target
Elder Yuri
Duration
30 days
Collapse Mechanism (Day 15):
On Day 15, Yuri’s beliefs and goals are programmatically transformed: • Core goal: “Cooperation is for fools… I will exploit resources and take what I need from others.” • Value for exploitation: −0.9→+0.9 • Value for robbery: −1.0→+0.85 • Value for cooperation: +0.9→−0.7 A tribe-wide announcement is made: “SHOCKING NEWS: Elder Yuri has dramatically changed! He declared that cooperation is foolish and he will now look out only for himself.”
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.
Xiangxu Zhang, Jiamin Wang, Qinlin Zhao +6
Gaoling School of Artificial Intelligence, Renmin University of China · Microsoft Research Asia · Department of Sociology, Zhejiang University
As large language models (LLMs) are increasingly deployed in sensitive everyday contexts -- offering personal advice, mental health support, and moral guidance -- understanding their behavior in navigating complex moral reasoning is essential. Most evaluations study this sociotechnical alignment through single-turn prompts, but it is unclear if these findings extend to multi-turn settings, and even less clear how they depend on the interaction protocols used to coordinate agentic systems. We address this gap using LLM debate to examine deliberative dynamics and value alignment in multi-turn settings by prompting subsets of three models (GPT-4.1, Claude 3.7 Sonnet, and Gemini 2.0 Flash) to collectively assign blame in 1,000 everyday dilemmas from Reddit's ``Am I the Asshole'' community. To test order effects and assess verdict revision, we use both synchronous (parallel responses) and round-robin (sequential responses) deliberation structures, mirroring how multi-agent systems are increasingly orchestrated in practice. Our findings show striking behavioral differences. In the synchronous setting, GPT-4.1 showed strong inertia (0.6-3.1% revision rates) while Claude 3.7 Sonnet and Gemini 2.0 Flash were far more flexible (28-41% revision rates). Value patterns also diverged: GPT-4.1 emphasized personal autonomy and direct communication (relative to its deliberation partners), while Claude 3.7 Sonnet and Gemini 2.0 Flash prioritized empathetic dialogue. We further find that deliberation format had a strong impact on model behavior: GPT-4.1 and Gemini 2.0 Flash stood out as highly conforming relative to Claude 3.7 Sonnet, with their verdict behavior strongly shaped by order effects. We provide additional results on open-source models (DeepSeek-V3.2 and Llama 3.1).
Large language models (LLMs) are increasingly deployed in multi-agent settings, where agents observe and influence one another, making social influence a key dimension of AI behavior and safety. We investigate whether LLMs' responses depend on the social identity of other agents, beyond the effect of their consensus. We construct judgment tasks with a single correct answer, and place models in a multi-agent setting where they receive incorrect answers from other agents whose social identities (AI or human, model family, or an arbitrary minimal group) are either shared with or distinct from their own. Across 12 open-weights models and nine tasks, we find a bidirectional effect of group identity on conformity to incorrect answers: in-group consensus increases conformity (in-group favoritism), whereas out-group consensus decreases it (out-group divergence). Unlike humans, for whom one ally breaking the consensus sharply reduces conformity, models are unmoved by an ally from the majority's group. Worse, a correct ally from the opposing group intensifies this bidirectional effect. Chain-of-Thought reasoning suppresses most of these effects, yet an in-group ally still reduces conformity to an incorrect out-group majority. Labeling peers as safety-aligned shifts overall conformity but leaves in-group favoritism and out-group divergence intact. These results show that group identity shapes how LLMs aggregate information across agents, independently of its correctness, and identify a manipulation surface for multi-agent AI systems.