Can LLM-driven agents mimic heterogeneous human behavior in GABMs?

Determine whether large language models, when used to drive agent decisions in generative agent-based models for epidemic modeling, can properly mimic the behavior of heterogeneous individuals across demographic and personality dimensions, specifically age, race, gender, and personality.

Background

Generative agent-based models (GABMs) replace hand-crafted behavioral rules with decisions produced by LLMs, aiming to better capture human decision-making during epidemics. While this approach can remove many simplifying assumptions, it raises concerns about whether LLMs can faithfully reproduce real human behaviors, especially across diverse populations.

The paper notes that LLM decisions can be biased and may not align with human behavior. A key unresolved issue is whether LLMs can accurately represent heterogeneity related to age, race, gender, and personality, which is crucial for realistic epidemic modeling and policy evaluation.

References

Furthermore, it is unknown if they could properly mimic the behavior of heterogeneous individuals in terms of age, race, gender, or personality.

— Generative Agent-Based Models for Complex Systems Research: a review  (2408.09175 - Lu et al., 2024) in Section 5: Epidemic modelling in the LLMs environment

Which of the two contributions dominates in a post-trained model is an open empirical question our design cannot settle.

— Readable, Faithful, Used: Three Dissociable Properties of Demographic Identity in a Language Model  (2608.18768 - Robbani, 19 Aug 2026) in Section 9, paragraph “A mechanistic candidate for the Sim2Real gap—and a warning about explaining it by representation quality”

Second, representativeness: we did not validate that synthetic agents’ self-generated preferences match those of the demographic groups they are sampled from, and it is unknown to what extent the underlying models have learned genuine correlations between demographics and aesthetic preference (the illustrative carpenter example in Section 3.2.2 should be read as a description of the mechanism, not as a validated association).

— FocusGen: Expanding Visual Design Exploration with a Simulated Focus Group of Persona Agents  (2608.28001 - Choi et al., 28 Aug 2026) in Section 6.4.1, “Synthetic Nature of Personas and the Missing Discriminative Test”

Smaller distances indicate closer population distributions, but leave open whether models assign answers to the right people or retain the right relationships.

— Artificial Societies Benchmark: A Validation Framework for Synthetic Research  (2609.30030 - Chidichimo et al., 24 Sep 2026) in Section 3, paragraph EV-1 Response distributions

Lower errors indicate closer agreement for the evaluated groups or people, while groups with insufficient human evidence remain unresolved.

— Artificial Societies Benchmark: A Validation Framework for Synthetic Research  (2609.30030 - Chidichimo et al., 24 Sep 2026) in Section 3, paragraph EV-2 Groups and individuals; Section 8, paragraph Representational fairness