SSP Forum: Huanxing Chen and Zeina Hashem (M.S. Candidates)

The
Symbolic Systems Forum
(community sessions of SYMSYS 280 - Symbolic Systems Research Seminar)
presents

Survey-Anchored Agents for Cross-Cultural LLM Simulation

Huanxing Chen (M.S. Candidate)
Symbolic Systems Program

and

Beyond Severity: Identifying Distinct Anxiety Driver Profiles in College Students Using Interpretable Machine Learning

Zeina Hashem (M.S. Candidate)
Symbolic Systems Program

Monday, May 4, 2026
12;30-1:20 pm PT
Computing and Data Science Building (CoDa), Room E160
In-person event, not recorded

Note: Lunch is provided, if pre-ordered, only for members of SYMSYS 280, but others are welcome to bring a lunch and eat during the presentation.

Abstracts:

Huanxing Chen, "Survey-Anchored Agents for Cross-Cultural LLM Simulation" (primary advisor: Michael Bernstein, Computer Science)
     (Direct Supervision by Jonne Kamphorst, now at Center for Socio-Political Data, SciencesPo) Large language models (LLMs) are increasingly used to simulate human populations for social science and policy research. However, growing evidence indicates that LLMs encode a systematic Western bias rooted in the composition of their training data. Applied to population simulation, this bias would be expected to produce both lower accuracy and artificially reduced diversity in the responses of simulated individuals from non-Western societies. Here, we first show that the standard approach — prompting LLMs with basic demographic characteristics such as age, gender, and education — indeed produces both distortions, and that both worsen as societies become more culturally distant from the United States. We then introduce survey-anchored agents, a new approach that conditions LLMs on representative survey responses rather than demographics alone. Using the World Values Survey, which covers 51 societies and 275 questions across 13 thematic domains, we compare survey-anchored agents against country and demographic baselines. Survey-anchored agents systematically achieve higher predictive accuracy and produce opinion diversity closer to that observed in human populations, with the largest gains in the most culturally distant countries. This work provides a scalable framework for building culturally robust generative agents, and suggests that anchoring artificial populations in representative surveys can help ensure that LLM-based research and applications generalize across diverse social contexts.

Zeina Hashem, "Beyond Severity: Identifying Distinct Anxiety Driver Profiles in College Students Using Interpretable Machine Learning" (primary advisor: Jamil Zaki, Psychology)
     Anxiety is highly prevalent among college students, yet interventions remain largely one-size-fits-all; partly because we lack a mechanistic understanding of what drives anxiety in this population, and whether those drivers differ across individuals. This study applies interpretable machine learning to survey data from 2,781 Stanford undergraduates to address both gaps simultaneously. An XGBoost model was trained to predict GAD anxiety scores from 12+ psychometric scales, deliberately excluding symptom-adjacent measures to surface informative, actionable predictors rather than circular correlates. SHAP values were computed to produce individual-level feature attribution vectors. Loneliness, self-compassion, subjective happiness, and stress emerged as the most robust predictors across five model specifications (R² = 0.46). Students were then clustered by their SHAP vectors, grouping people by the mechanism driving their anxiety, yielding four stable, interpretable profiles (bootstrap ARI = 0.997). Substantial GAD variance within clusters revealed that anxiety driver profiles and anxiety severity are dissociable: students sharing risk mechanism profiles do not necessarily share a symptom burden. These findings suggest that anxiety in college students reflects meaningfully distinct vulnerability pathways, and these profiles may offer a more actionable basis for personalized campus intervention than severity-based assessment alone.