Optimizing auxiliary agent attributes beyond exogenous variables in LLM-powered simulations

Develop an optimization framework for selecting and endowing additional agent attributes—such as demographics, personalities, and other traits—beyond the structural causal model’s exogenous variables for large language model–powered agents, so as to improve simulation fidelity while avoiding redundancy and unintended interactions.

Background

The authors adopt a minimalist approach to agent construction, endowing LLM-powered agents primarily with goals, constraints, roles, names, and the exogenous variables required by the structural causal model. They note that additional attributes (e.g., demographics, personality traits) might enhance fidelity but could introduce redundancy or unforeseen interactions.

They explicitly state that optimizing which auxiliary attributes to include is unclear, motivating the need for a principled method to balance realism and experimental control in automated social-science simulations with LLM agents.

References

However, it is unclear how to optimize this process.

— Automated Social Science: Language Models as Scientist and Subjects  (2404.11794 - Manning et al., 2024) in Subsection “Future research,” Section “Conclusion”

The composition of a persona is of particular note; it is not clear which attributes should make up a persona and whether this selection should be determined based on an a priori theory, constructed in a data-driven manner, or a combination of both.

— Total Simulated Survey Error: Designing and Diagnosing Survey Responses from Large Language Models  (2609.10280 - Sen et al., 9 Sep 2026) in Section 3.1, subsection “Persona Construction”

Two directions we have not explored are: (i) personality-conditioned harnesses~\citep{chi2024amongagents} that condition the agent's chat and plan style on an assigned imposter or crewmate persona, which could decouple deception strategy from the underlying VLM's default style; and (ii) meta-harness optimization~\citep{lee2026metaharness} that searches over the harness implementation itself rather than the five hand-specified axes. Both are natural follow-ups to the present per-axis ablation and we leave them to future work.

— Lies We Can See: Joint Verbal and Non-Verbal Deception by VLM Agents in Embodied Social Interactions  (2608.30428 - Ahn et al., 31 Aug 2026) in Limitations, subsection “Agent harness scope”

Future work should investigate when behavioural associations transfer across tasks and models, and whether combining multiple empirically grounded individual characteristics can reproduce richer forms of human behavioural diversity.

— Conditioning LLMs on Social Value Orientation improves behavioural alignment in a sequential social dilemma  (2610.01667 - Saponara et al., 1 Oct 2026) in Section Conclusions

The life simulator writes English conversations, and the lives it writes follow the generating LLMs' notion of how a typical life unfolds. Associations that are specific to a culture, a community or an individual may therefore be learned less well, and extending the simulator to other languages and to more diverse lives is left to future work.

— Madeleine: Learning Involuntary Recall for Conversational Memory from Simulated Lives  (2610.01118 - Shi, 1 Oct 2026) in Limitations, “Coverage of the simulated lives”

This process also produces two other traits: hostile non-cooperation and mental and physical impairment. However, due to their complexity, high stakes, and relative rarity, we leave the simulation of these traits to future work.

— InterviewPlayground: A Simulation Environment for Evaluating AI Interviewers  (2610.12023 - Ivey et al., 8 Oct 2026) in Section 3.1, page 3