AI Agents Alone Are Not (Yet) Sufficient for Social Simulation

1 Introduction

• LLM-agent societies are used to study collective outcomes such as diffusion, polarization, market behavior, and policy interventions, with scalability and repeatability as key attractions.These settings can make human studies costly, slow, or ethically constrained.Source: §1 · p. 1
• Human-like demonstrations can encourage the inference that role-playing fidelity and network placement are sufficient for society-like collective validity.The paper cautions that transcript plausibility does not establish how information, incentives, and constraints generate or propagate.Source: §1 · p. 1
• LLM-agent pipelines prioritize role consistency and conversational plausibility, while social simulation requires faithful decision processes and specified generative mechanisms.The authors propose an environment-involved formulation and practical actions to improve reliability, interpretability, and epistemic clarity.Source: §1 · p. 2

2.1 Scope

• The paper scopes its analysis to LLM-integrated agents used as human proxies in social simulations with collective-level interpretation, rather than task performance or predictive accuracy.It does not aim to survey all social simulation methods or non-agent-based approaches.Source: §2.1 · p. 2

2.2 Current Approaches

• Current practice commonly assumes plausible individual agents will yield realistic collective dynamics, and is organized around individual- and collective-level approaches.Individual-level work configures human proxies; collective-level work studies how populations are composed, connected, and interacted.Source: §2.2 · p. 2

2.2.1 Individual-level Simulation

• Individual-level methods use prompts to shape personas or training procedures to ground, fine-tune, or reinforce agent behavior.These methods remain behaviorist and can struggle with deeper motivations, long-term stability, and human decision-making fidelity [36].Source: §2.2.1 · pp. 2–3

2.2.2 Collective-level Simulation

• Collective-level approaches include controlled interaction scenarios and heterogeneous populations designed to reproduce system-level dynamics.Examples range from dialogue, collaboration, and goal-centred tasks to economic, political, and online-platform simulations.Source: §2.2.2 · p. 3

3.1 Fundamental Mismatches

• Role-playing plausibility can diverge from faithful human simulation because coherent text does not ensure decisions reflect stable preferences, constraints, incentives, or social learning.Prompt sensitivity and training signals that reward imitation rather than decision-making under constraints can leave agents fragile under counterfactual interventions.Source: §3.1 · pp. 3–4
• Social outcomes often arise through agent–environment co-dynamics, so agent messaging alone can miss how institutions and platforms shape observation, action, and incentives.Omitting these mechanisms can misattribute effects to social behavior or misrepresent interventions that change exposure, incentives, or constraints.Source: §3.1 · p. 4

3.2 Current Gaps

• The paper distinguishes conceptual mismatches from concrete pipeline gaps that leave assumptions insufficiently operationalized or outcomes insufficiently tested for robustness.The mismatch lens contrasts proxy objectives with epistemic goals; gaps concern missing or underdeveloped pipeline components.Source: §3.1 · p. 3, §3.2 · p. 5

3.2.1 Evaluation of Simulation Results

• Evaluation remains fragmented, with limited quantification of run-to-run variance and baseline comparison, while output-centric measures can reward plausibility without behavioral validity.Collective evaluation may use human-likeness judgments or alignment to aggregate traces, and weak validation risks selective reporting and brittle policy recommendations.Source: §3.2.1 · p. 5

3.2.2 Interaction Dynamics and State Update

• Interaction designs range from ambiguous dialogue-first exchanges to explicit rule-mediated updates, but both leave important modeling trade-offs.Dialogue can make trajectories sensitive to phrasing and turn order; rules improve clarity but may fail to represent real-world heterogeneity and diverse decision processes.Source: §3.2.2 · p. 6
• Scheduling and synchronization are part of the model because sequential execution can differ from asynchronous social activity and alter exposure and feedback loops.Implementations must specify concurrent interactions, interleaving, and what agents observe when deciding.Source: §3.2.2 · p. 6

3.2.3 Initialization and Information Priors

• Initialization choices about population composition, network structure, scale, and who knows what can encode strong assumptions and dominate downstream dynamics.Coarse demographic matching may miss attribute correlations and latent variables; prompt-level priors can conflate information availability with awareness, while reduced scale may distort nonlinear phenomena.Source: §3.2.3 · pp. 6–7

Formulation

• The formulation represents social simulation with explicit agent, environment, network, context, visibility, scheduling, transition, and initial-condition components.It is defined for a fixed number of agents N and simulation horizon T.Source: §4 · p. 8
• Separating agent generation from environment mechanisms makes information exposure, action timing, and institutional consequences independently inspectable.The formulation supports controlled policy changes through environment state and mechanisms, but does not guarantee valid simulation.Source: §4 · p. 8

Action 1: Treat the environment as a first-class, auditable object

• Treat the environment as an auditable object by versioning and logging its state and mechanisms so auditors can reconstruct constraints, incentives, and information frictions.When manual specification is infeasible, an empirical dynamics model may be used if intermediate states and uncertainty diagnostics remain exposed.Source: §4 · p. 8

Action 2: Move evaluation beyond plausibility to mechanistic and counterfactual reliability

• Evaluate whether decisions respond to constraints and incentives, targeted mechanism edits produce stable directional changes, and comparative conclusions survive irrelevant implementation changes.The paper recommends ablations, negative controls, and one-component-at-a-time counterfactual sweeps, reporting distributions rather than selected trajectories.Source: §4 · pp. 8–9

Action 3: Interpret simulation outcomes with epistemic caution and explicit uncertainty

• Interpret outputs as conditional implications of assumed mechanisms and implementation choices, not automatic evidence of social dynamics.Report uncertainty and sensitivity to initial conditions, visibility, scheduling, transitions, and the LLM configuration.Source: §4 · p. 9

Individual-fidelity-sufficient thesis

• The individual-fidelity-sufficient thesis holds that human-aligned agents could yield realistic collective behavior without explicit environment modeling; supporting work uses demographic personas [4] [49] [83] or deeper cognitive substrates [36].The authors argue that agent-level refinements do not specify information disclosure, incentives, or enforcement, so role-consistent agents may still produce externally implausible dynamics [33] [80].Source: §5 · p. 9

Rigidity worry

• The rigidity worry says explicit rules may revive rigid classical models, but omitting structure relocates assumptions into prompts, schedules, and interfaces, making them harder to audit.The authors advocate making key mechanisms explicit, even approximately, rather than requiring full environment specification.Source: §5 · p. 9

Conclusion

• LLM expressiveness alone does not confer epistemic validity; simulations must make assumptions about interaction, information exposure, and institutional coupling explicit and testable.Otherwise, simulations risk becoming persuasive narratives rather than scientific instruments.Source: §6 · p. 9
• The paper calls for mechanism-level audits, treating environments as first-class objects, separating proxy objectives from epistemic goals, and foregrounding uncertainty.It argues that aggregate human-correlation is insufficient evidence.Source: §6 · p. 9
• Within the paper's stated scope, these principles are presented as necessary for moving toward reliable scientific understanding and policy reasoning rather than fragile, overconfident conclusions.Source: §6 · p. 9

• Social simulation models individual cognition, decision-making, and interaction to generate, explain, or predict collective social phenomena.It supports controlled, repeatable, and counterfactual analysis of complex systems and policy interventions.Source: A More Details of Our Scope · p. 16
• LLM-integrated agents are studied as flexible human proxies, against earlier rule-based approaches that can impose stylized behavioral assumptions and require manual redesign across settings.LLMs are viewed as enabling open-ended communication and role-conditioned behavior.Source: A More Details of Our Scope · p. 16
• The paper focuses on methodological assumptions of agent-based social simulations interpreted at the collective level, not task performance, predictive accuracy, or a comprehensive survey.It excludes non-agent-based approaches, LLM personality and interpretability, and agent applications outside social simulation.Source: A More Details of Our Scope · p. 16