Beyond Demographics - How AI Can Better Understand Human Values Through Psychological Simulation

Paper info: Haijiang Liu, Qiyuan Li, Chao Gao, Yong Cao, Xiangyu Xu, Xun Wu, Daniel Hershcovich, Jinguang Gu. 2025. Beyond Demographics: Enhancing Cultural Value Survey Simulation with Multi-Stage Personality-Driven Cognitive Reasoning. The 2025 Conference on Empirical Methods in Natural Language Processing.

When we try to predict how someone will respond to questions about their deepest values—whether family is important, how much they trust others, or what they prioritize in life—what information do we really need? Most AI systems today rely heavily on basic demographic data: age, gender, education level, income. But as anyone who has met people from diverse backgrounds knows, demographics alone paint an incomplete picture of human complexity.

A new research paper accepted at EMNLP 2025 challenges this demographic-centric approach with a fundamentally different strategy: simulating the actual cognitive processes that drive human decision-making. The work, titled “Beyond Demographics: Enhancing Cultural Value Survey Simulation with Multi-Stage Personality-Driven Cognitive Reasoning,” introduces MARK (Multi-stAge Reasoning frameworK), which achieves a notable 10% improvement in predicting human survey responses by modeling personality-driven cognition rather than relying solely on demographic stereotypes.

The Problem with Demographics-Only Approaches

Current approaches to simulating human survey responses face a critical limitation: they tend to produce stereotypical responses that fail to capture the nuanced decision-making processes characteristic of real human behavior. When researchers prompt large language models with demographic information like “25-year-old married female with two children,” the models often fall back on broad generalizations rather than modeling the complex interplay of stress, personality, and situational factors that actually drive human responses.

The authors demonstrate this limitation through experiments on the World Values Survey, where traditional demographic prompting approaches showed significant gaps in both accuracy and distribution alignment with actual human responses. More concerning, these approaches often reinforced demographic stereotypes rather than capturing genuine individual differences in values and decision-making.

Enter MARK: Simulating the Mind, Not Just the Demographics

MARK takes a radically different approach by modeling the cognitive architecture underlying human decision-making. Drawing from MBTI personality type dynamics theory, the framework simulates how different personality types process information and make decisions under varying levels of stress.

The key insight driving MARK is that human responses to survey questions are significantly influenced by two factors often ignored by demographic approaches:

Mental status and stress levels - How life circumstances create psychological pressure
Cognitive processing patterns - How personality types systematically differ in information processing

How MARK Works: A Four-Stage Cognitive Simulation

MARK operates through four interconnected stages that progressively build a psychological profile and simulate decision-making:

Stage 1: Stress Analysis

Rather than treating demographic features as static descriptors, MARK evaluates how each aspect of a person’s life situation contributes to their overall stress level. A 25-year-old with two children might experience moderate stress from financial pressures, while their urban environment and full-time work add additional layers of complexity. The system assigns stress scores (0-100) to each demographic feature and uses these to generate more nuanced sociodemographic prompts.

Stage 2: Personality Prediction

Using the stress-analyzed demographic profile, MARK predicts cognitive functions based on MBTI personality theory. Rather than directly mapping demographics to personality (which would be problematic), the framework treats demographic features as environmental pressures that may shape cognitive function preferences over time. The system predicts the four cognitive functions in their hierarchical order: Dominant, Auxiliary, Tertiary, and Inferior.

Stage 3: Cognitive Reasoning

This stage simulates how each cognitive function would process the specific survey question under the person’s current stress level. Different cognitive functions—like Extraverted Thinking (logical, goal-oriented) versus Introverted Feeling (values-based, empathetic)—systematically approach the same question differently. The system generates responses for each function and assigns confidence weights based on the cognitive hierarchy and stress impact.

Stage 4: Synthesis

Finally, MARK combines the outputs from all cognitive processes, weighing them according to their position in the personality hierarchy and stress-influenced activation patterns. This produces both a final survey response and an interpretable explanation of the decision-making process.

Impressive Results Across Cultures

Testing on the World Values Survey data from both US and Chinese populations, MARK demonstrated substantial improvements:

10% higher accuracy compared to existing baselines
Better distributional alignment with actual human response patterns
Robust performance even with uncertain personality predictions
Cross-cultural generalization maintaining effectiveness across Eastern and Western cultural contexts

The framework proved particularly effective for certain demographic clusters. For instance, it showed strong performance predicting responses from married individuals working full-time jobs and skilled workers in private business—groups where the interplay of work stress, family responsibilities, and cognitive processing patterns creates complex decision-making scenarios.

Critical Considerations and Limitations

While MARK represents a significant advance, the authors acknowledge several important limitations that deserve attention:

Psychological Validity vs. Computational Tractability: The framework relies on MBTI theory, which has well-documented limitations in psychological research compared to empirically validated models like the Big Five. The authors chose MBTI primarily because its hierarchical cognitive function structure provides explicit rules for multi-step reasoning, while the Big Five lacks such systematic frameworks for modeling cognitive processes.

Personality Prediction Challenges: The system must infer personality from demographic information, which introduces potential bias. While the approach is more sophisticated than direct demographic-to-attitude mapping, it still requires making assumptions about how environmental factors shape personality development.

Cultural and Linguistic Coverage: Testing was limited to US and Chinese populations using English and Chinese prompts, which may not capture the full diversity of global cultural approaches to values and decision-making.

MARK’s success suggests several important directions for the field:

Moving Beyond Demographics: Rather than relying on demographic categories that can reinforce stereotypes, AI systems can model the psychological processes that actually drive human behavior.
Interpretable AI for Social Research: By simulating cognitive processes, MARK provides interpretable explanations for its predictions—crucial for social scientists seeking to understand rather than just predict human behavior.
Stress-Informed Modeling: The framework’s emphasis on how stress influences cognitive functioning offers a more nuanced approach to understanding human decision-making under pressure.
Cross-Cultural Robustness: The consistent performance across US and Chinese contexts suggests that cognitive process modeling may be more universally applicable than demographic stereotyping.

Looking Forward

This research opens compelling avenues for leveraging AI in social science research while reducing reliance on costly human experiments. The framework’s emphasis on modeling psychological processes rather than demographic categories offers a path toward more accurate and less biased simulation of human preferences.

However, the work also highlights the ongoing challenge of balancing psychological realism with computational feasibility. Future research might explore integrating more empirically validated personality frameworks or incorporating additional mediating variables to further reduce bias risks.

As AI systems increasingly attempt to model and predict human behavior, MARK demonstrates the value of grounding these efforts in cognitive and psychological theory rather than surface-level demographic patterns. The result is not just better accuracy, but more interpretable and potentially less biased AI systems that can genuinely contribute to our understanding of human values and decision-making.

You can find the full paper on arXiv.

Citation:

@misc{liu2025demographicsenhancingculturalvalue,
      title={Beyond Demographics: Enhancing Cultural Value Survey Simulation with Multi-Stage Personality-Driven Cognitive Reasoning}, 
      author={Haijiang Liu and Qiyuan Li and Chao Gao and Yong Cao and Xiangyu Xu and Xun Wu and Daniel Hershcovich and Jinguang Gu},
      year={2025},
      eprint={2508.17855},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.17855}, 
}