Designing agents that accurately simulate human behavior and social interactions

Develop autonomous agents capable of simulating realistic and comprehensive human behavior and social interactions within world models to support faithful social simulation and analysis.

Background

The survey argues that world models must incorporate not only physical dynamics but also social dimensions to be useful in many real-world scenarios, such as urban environments and social computing systems. While LLM-based agents provide new opportunities to simulate human-like behaviors, achieving realism and comprehensive coverage of social interactions remains challenging.

In discussing future directions, the authors explicitly identify the design of autonomous agents that can authentically replicate human behavior and social interactions as an unresolved issue, highlighting the current limitations of agentic workflows and the need for more principled approaches informed by theories of human behavior and cognition.

References

However, designing autonomous agents that can simulate realistic and comprehensive human behavior and social interactions remains an open problem.

— Understanding World or Predicting Future? A Comprehensive Survey of World Models  (2411.14499 - Ding et al., 2024) in Section 6.2 Enriching the Social Dimension (Open Problems and Future Directions)

Thus, designing LLMs specialized for social simulations is an open interdisciplinary engineering challenge, requiring input from social scientists as much as LLM designers; there is also the question of dual use --- how would such specialized models, which might have different types of safety guardrails compared to existing assistive LLMs, be deployed and made available to the research community without being misused?

— Total Simulated Survey Error: Designing and Diagnosing Survey Responses from Large Language Models  (2609.10280 - Sen et al., 9 Sep 2026) in Section 6, “Discussion,” subsection “Using TS2E”

Consequently, achieving joint controllability over macro trajectories and micro pedestrian motions in photorealistic rendering remains an open challenge.

— Controllable and Photorealistic Pedestrian Risky Motion Generation for End-to-End Driving Safety Evaluation  (2610.06171 - Liu et al., 5 Oct 2026) in Section II-A, “Safety-Critical Scenario Generation”