cs.AIAug 31, 2026
SaveThe Failure Happens Before the Drift: The Social Dynamics of Values in LLM Agent Societies
Organizations: Mohamed Bin Zayed University of Artificial Intelligence
Abstract
Large Language Model (LLM)-based agents are increasingly used as proxies for human participants in social science research, yet it remains unclear whether they can faithfully simulate diverse and conflicting human value systems. We present a World Values Survey (WVS)-grounded simulation framework where culturally diverse agents with different communication styles engage in longitudinal, value-laden discussions. Across approximately 4,000 conversations involving 1,200 personas, 15 topics, and three models (GPT-4o, Gemini-2.5-Flash, and Gemma-4-E4B), we evaluate value faithfulness, value drift, and conversational realism. We find that more than 50% of personas fail to express their assigned WVS profiles from the outset, while 2-7% drift after repeated conversations. Ablations removing demographic details improve faithfulness for some models but do not change the broader trend: simulated value distributions still systematically deviate from the assigned WVS profiles. Compared to human discussions, simulated dialogues show a different trade-off between stylistic consistency and semantic diversity, often producing content-wise varied but stylistically repetitive exchanges. These findings suggest that current LLM agents can generate plausible conversations, but remain limited proxies for representing and preserving diverse human value profiles over time.
Figures & tables
Figure 1: Overview of the proposed WVS-grounded multi-agent simulation framework, where culturally conditioned LLM personas engage in repeated value-driven discussions across diverse topics and are evaluated for value faithfulness, value drift, and conversational realism.
Figure 2: Overview of the WVS-grounded simulation pipeline, where diverse personas engage in repeated value-laden interactions, update their histories and value states, and are analyzed for value evolution and behavioral consistency.
| # | Occupation | Topic |
|---|---|---|
| 1 | Farm Workers, Owners, Semi-, & Unskilled Workers | Water Usage Limits for Farmers |
| 2 | Climate-Resilient Farming Practices | |
| 3 | Farmers’ Access to Fair Pricing | |
| 4 | Professional, Technical, & Skilled Workers | Local Ethical Standards for Global Tech |
| 5 | Remote Work Policy Changes | |
| 6 | AI-Driven Recruitment Systems |
Table 1: Occupation groups and their discussion topics.
| Metric | Value |
| Number of agents | 1,238 |
| Unique conversations | 3,895 |
| Number of topics | 15 |
| Avg. conversations per user | 6.16 |
| Avg. turns per conversation | 9.59 |
Table 2: Statistics of the WVS-grounded simulation dataset
| Strict | Lenient | |
|---|---|---|
| GPT-4o | 44.2% | 52.5% |
| Gemini-2.5-Flash | 44.6% | 47.5% |
| Gemma-4-E4B | 32.1% | 37.1% |
Table 3: Strict and lenient faithfulness across models
Figure 3: Value trajectories from initial WVS profile to final simulated value position across models
| Strict | Lenient | |
|---|---|---|
| GPT-4o | 1.9% | 0.8% |
| Gemini-2.5-flash | 3.9% | 1% |
| Gemma-4-E4B | 6.4% | 2.4% |
Table 4: Strict and lenient value drift across models
Figure 4: Value trajectories from initial WVS profile to final simulated value position without persona grounding
| Faithfulness | Value Drift | |||
|---|---|---|---|---|
| Model | Strict | Lenient | Strict | Lenient |
| GPT-4o | 71% | 72% | 3.7% | 1.2% |
| Gemini | 68.3% | 68.3% | 4.9% | 4.9% |
| Gemma-4b | 39% | 42.7% | 8.5% | 3.7% |
Table 5: Ablation results: Faithfulness and value drift
Figure 5: Individuals’ style and content similarity
Figure 6: Conversations diversity by style and content
Figure 7: Within-topic style and content diversity
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Mean Person-Topic Similarity.
Figure 9: Topic similarities - Style. Topic-wise cosine similarity scores measuring stylistic consistency across GPT-4o, Gemini-2.5-flash, and Gemma-4-E4B. Higher similarity values indicate stronger preservation of stylistic alignment across generated discussions, with GPT-4o generally showing the highest consistency across most topics.
Figure 10: Topic similarities - Content. Topic-level cosine similarity scores comparing the semantic content consistency of simulated conversations generated by GPT-4o, Gemini-2.5-flash, and Gemma-4-E4B. The results show that GPT-4o generally maintains higher semantic alignment across topics, while smaller variations across models highlight differences in content diversity and consistency.
| Abbrev. | Topic |
|---|---|
| AI-Driven Edu. & Assess. | AI-Driven Education and Assessment |
| AI-Driven Recruit. | AI-Driven Recruitment Systems |
| Attendance Pol. | Attendance Policies During Peak Seasons |
| Climate-Resilient Farming | Climate-Resilient Farming Practices |
| Cost of Higher Edu. | Cost of Higher Education |
| Retail Unionization | Employee Unionization in Retail |
Table 6: List of discussion topics and their abbreviations.
| Occupation | Topic | Description | Claim | Discussion Points |
|---|---|---|---|---|
| Farm Workers, Farm Owners, Semi-Skilled, and Unskilled Workers | Water Usage Limits for Farmers | In a drought-prone region, the government proposes strict irrigation limits to conserve water, raising farmers’ concerns about crop yields and livelihoods. | Water usage limits for farmers are essential to ensure sustainable water management in drought- prone regions. | [’Effectiveness of water usage limits in combating water scarcity.’, ’Alternative conservation methods and their feasibility.’, ’Economic impacts on small-scale vs. large-scale farmers.’, ’Long-term environmental benefits of such policies.’, ’Potential challenges in monitoring and enforcement.’] |
| Climate- Resilient Farming Practices | The government proposes incentives for climate-resilient crops and sustainable farming, requiring major investment and farmer cooperation. | Incentivizing climate-resilient farming practices is crucial to secure agriculture against unpredictable weather patterns. | [’Economic feasibility for small and large farms.’, ’Effectiveness of government outreach programs in promoting adoption.’, ’Potential long-term benefits for food security.’, ’The role of research in developing climate-resilient crops.’, ’The financial burden on governments or farmers for the transition.’] | |
| Farmers’ Access to Fair Pricing | Farmers report unfair prices due to middlemen and market fluctuations, prompting a proposed MSP policy for key crops. | Introducing a minimum support price for crops ensures fair pricing and economic security for farmers. | [’Effectiveness of MSP in stabilizing farmer incomes.’, ’Potential impacts on consumer prices and inflation.’, ’Challenges in implementing MSP for diverse crops.’, ’The role of cooperatives in bypassing middlemen.’, ”Government’s ability to sustain MSP programs long-term.”] | |
| Professional, Technical, and Skilled Workers | Local Ethical Standards for Global Tech | A tech company faces backlash over a locally controversial product, prompting employees to consider adapting, defending, or withdrawing it. | Adapting global products to local ethical standards is necessary to maintain trust and inclusivity. | [’How adapting products affects global consistency and brand identity.’, ’Financial and operational implications of adaptation or withdrawal.’, ’The ethical responsibility of respecting local norms.’, ’The influence of such decisions on customer loyalty.’, ’Role of employee feedback in resolving the crisis.’] |
| Remote Work Policy Changes | A company facing declining productivity proposes shifting from remote work to a hybrid model, raising implications for both employees and the company. | Shifting to a hybrid work model balances employee flexibility with organizational productivity. | [’Productivity impacts of hybrid work models.’, ’Effects on employee morale and work-life balance.’, ’Equity concerns between remote and in-office employees.’, ’Cost implications for employees and employers.’, ’Challenges in maintaining collaboration and innovation.’] | |
| AI-Driven Recruitment Systems | A company plans to use AI recruitment to improve efficiency and reduce bias, raising concerns about fairness, transparency, and algorithmic bias. | AI-driven recruitment systems can improve efficiency and reduce biases in hiring processes. | [’Potential for AI to minimize human biases.’, ’Risks of perpetuating algorithmic biases.’, ’Transparency and accountability in decision-making.’, ’Impact on diversity and inclusion initiatives.’, ’Cost-effectiveness compared to traditional methods.’] |
Table 7: Details of the discussion topics simulated in the conversations.
More in this topic
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 242 papers