acceptodds
Under review as a conference paper at ICLR 2027

When Do LLM Agents Reproduce Human Social Structure? An Identifiability Audit of Generative Social Simulators

Abstract

Large language models (LLMs) are increasingly used as agents to simulate human societies, and such simulators are often evaluated by whether the social structures they generate resemble real ones. However, realistic-looking social structure does not necessarily imply realistic agent behavior: the observed structure may arise from simulator architecture or even from superficial features of the prompt. We introduce the Simulator Identifiability Audit, a framework for testing whether emergent social structure is actually attributable to an LLM agent policy. The audit compares each simulator with an architecture-matched random twin that replaces LLM choices by chance while holding the rest of the simulator fixed, evaluates both at matched edge counts, and permutes labels to separate effects of simulated individuals from effects of their representations. Across six social-simulation architectures, real and synthetic social networks, three commercial LLM providers, and an open-weight model, we find three consistent patterns. First, a minimal persona-agent architecture distilled from recurring design choices in published simulators is structurally silent: none of its evaluated network properties differs detectably from its random twin. Second, exposing peer traits makes the policy detectable, improving clustering through convergent selection of the same trait-attractive peers rather than homophily, while degrading degree structure. Third, apparent successes can be artifacts: memory-driven community formation disappears at matched edge counts, while names and list positions systematically bias partner selection, with female-associated names selected up to 8.7x as often as male-associated names in a synthetic-name probe.

Primary area:
datasets and benchmarks
Keywords:
large language model agents, generative agents, social simulation, social network generation, evaluation, identifiability, benchmarking, presentation bias

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_cWS6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether the similarity between LLM-generated social networks and real human social networks comes from the LLM agents' decision-making policies, or from the simulator architecture and input representations. To address this question, the authors propose the Simulator Identifiability Audit, which uses architecture-matched Random Twins, Matched Edge Counts, and Label Permutation to evaluate generative social simulators in three stages: policy detectability, structural fidelity improvement, and attribution to individuals. The authors conduct experiments across six simulator architectures, real and synthetic social networks, and multiple LLMs. They also release the ClassNetBench benchmark.

Critical strengths

The research question is quite novel. Existing studies on LLM-based social simulation rarely distinguish the contributions of LLM policies, simulator architectures, and input representations. This paper clearly identifies this attribution problem and proposes an architecture-matched Random Twin and a three-stage audit framework to address it.

Critical weaknesses
  1. The authors project generated networks to a common edge count (m∗m^*) to reduce the influence of network size differences on structural metrics. This is reasonable when comparing connection patterns at a fixed edge count. However, the number of edges can also be an important outcome influenced by the LLM policy itself. The paper uses the minimum edge count across generated networks as m∗m^*, while keeping the observed network unchanged. When the LLM generates far fewer edges than the real network, this projection may limit the higher-order structures that can be observed. Therefore, the paper should more clearly distinguish between the overall effect of the policy on the generated network and its effect on network structure after controlling for the edge count.

  2. The experiments mainly build on a shared framework of candidate selection and edge accumulation, with additional mechanisms such as information exposure, choice volume, memory, and network backbones. These tasks are essentially structured social partner selection, which is still quite different from free-form conversations, two-way interactions, dynamic relationship development, and long-term planning in many real-world generative social simulation systems.

Questions for authors
  1. The paper mentions that each architecture is calibrated using its Random Twin, and that the Random Twin is calibrated to the same structural targets used in evaluation. Which network statistics are used during calibration? Do these statistics overlap with the six metrics used to evaluate structural improvements? Are the same real networks used for both calibration and final evaluation?

  2. Does the disappearance of T3's community improvement really show that the improvement is only caused by differences in edge counts, or could the current matching method be controlling away part of the effect of memory? Across different datasets, what proportion of the observed network's edges does the common edge count (m∗m^*) represent for T3? How sensitive are the results to the choice of (m∗m^*)?

Reviewer_dZCvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work examines whether the social structures generated by LLM agents are actually a result of their own policies, or whether they are largely shaped by superficial factors such as simulator architecture and prompting. It introduces the Simulator Identifiability Audit, which uses architecture-matched random twins as controls, together with edge-count matching and label permutation, to disentangle these effects. The results suggest that much of the observed social structure is an artifact of biases and edge-count mismatches, rather than evidence of agent strategies. Exposing peer traits improves individual-level clustering, but degrades other structural properties, such as the degree distribution.

Critical strengths
  1. The work establishes a useful evaluation framework for distinguishing the sources of network structure. By combining random controls, edge-count matching, and label permutation, it separates the effects of agent policies from other factors and provides a reusable procedure for evaluating social structure simulation.

  2. The evidence is fairly thorough. For example, edge-count matching eliminates the apparent modularity advantage of memory mechanisms. Identifying and ruling out such apparent improvements shows that The work takes the evaluation seriously rather than focusing only on positive results.

Critical weaknesses
  1. Some key analytical decisions were made after observing the experimental results, which weakens the reliability of the conclusions. Both the use of edge-count matching as the primary analysis and the choice of the equivalence threshold were decided post hoc. Although The work includes additional density-sensitivity analyses, further evidence is needed to make the conclusions more convincing.

  2. The external validity is limited. In some experimental conditions, the number of edges after matching is far below that of the corresponding real networks, making the resulting metrics difficult to interpret. In addition, the range of real networks tested and the number of runs with open-weight models are limited, making it difficult to draw broadly generalizable conclusions.

Questions for authors
  1. Since the primary comparison and equivalence margins were determined after observing the results, how sensitive are the main conclusions to alternative edge-matching procedures and equivalence margins?

  2. Could the audit be evaluated on a broader range of real-world networks and more dynamic social simulation architectures, using appropriate architecture-matched random twins?

Reviewer_Ag4UOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Simulator Identifiability Audit, a framework for testing whether simulated social structure is actually attributable to an LLM agent policy, rather than to the simulator architecture or to superficial features of the prompt. The audit has three devices: an architecture-matched random twin that replaces the LLM's choices with random choices while holding the rest of the simulator fixed; evaluation of the LLM simulator and its twin at matched edge counts; and label permutation, which separates effects of the simulated individuals from effects of the names and list positions used to display them.

Across six social-simulation architectures, real and synthetic social networks, three commercial LLM providers, and an open-weight model, the authors claim three consistent patterns. First, a minimal persona-agent architecture distilled from published simulators (T0) is structurally silent: none of its evaluated network properties differs detectably from its random twin. Second, exposing peer traits (T2/T5) makes the policy detectable; clustering improves through convergent selection of the same trait-attractive peers rather than through homophily, while degree structure degrades. Third, apparent successes can be artifacts: memory-driven community formation (T3) disappears at matched edge counts, and names and list positions systematically bias partner selection.

The interesting part of the paper is that it proposes a new way to compare a simulated social network with the real one. Instead of asking only whether the generated structure looks realistic, it asks whether the LLM policy produced that structure, component by component, and whether any improvement follows the simulated persons or merely their labels. The authors also release ClassNetBench, the software that implements these comparisons, together with the pipeline for the public datasets.

Critical strengths
  1. The comparison method changes a substantive conclusion. For T3, the observed-density projection gives a modularity-error advantage of −0.102 over the twin, alongside a 56% edge shortfall for the LLM versus 5% for its twin. At matched edge count the difference becomes +0.008 (90% CI [−0.004, 0.020]); the density sweep and seed-wise matching support the same interpretation. The authors explicitly withdraw the predicted community benefit of memory and disclose when the primary analysis changed. This is persuasive evidence that the proposed way of comparing simulated and real networks matters in practice.

  2. The attribution controls are complementary and well documented, and the software makes them reusable. Under T0, popularity consistency remains in displayed-label space (ρ=.78) but disappears in real-student space (.01); under T2/T5 it instead follows persons. The clustering improvements survive the decision-level shuffle twin. The forced-full-scoring experiments separate the sorting instruction's score gradient from selection behavior under a truncated response budget. Verbatim prompts, incident disclosures, public-data replications, planted-mechanism checks, and the ClassNetBench release make the audit inspectable and applicable to other simulators.

Critical weaknesses
  1. Statements such as "the LLM has no effect on structure" are broader than what was actually tested. The paper already identifies T0 as a lower bound, distinguishes error-space effects from raw-statistic changes, and limits inference to the observed sites. These qualifications are appropriate. However, P1–P3 do not imply equality of all six component errors; ego-attribute and label effects remain possible. For T0, the degree and clustering errors are undetermined, and only the remaining four mean errors satisfy the chosen equivalence margins. Those margins were selected after the initial analysis, and some cells have very sparse projections (e.g., T1/B: 19 of 175 edges). The existing sensitivity checks are reassuring, but the strongest statements in the abstract and discussion should consistently describe conditional, error-space results and should not treat non-detection as behavioral or structural equivalence.

  2. Numerical consistency needs checking. Table 4 defines Δtwin = S − U, yet its T0 coupling entries give .035 − .051 = −.016 rather than the reported −.013, and the assortativity entries give −.022 − .004 = −.026 rather than −.019. These differences exceed rounding and may reflect a reporting error or changed projection conditions. Relatedly, fixed-volume twins initially use uniform scores, which can affect edge increments under score-based update rules; the empirical-score check preserves 17 of 18 in-vivo verdicts, but T5's isolation degradation becomes non-significant and should not be presented as equally robust. The headline "up to 8.7×" name effect should also specify L=30, since Appendix I reports 9.8× at L=10. These appear fixable, but numerical consistency matters especially for a paper whose contribution is an audit.

Questions for authors
  1. Can you reconcile the Table 4 arithmetic, stating whether the LLM graph and m* are identical in its U, S, and Δtwin comparisons? Please provide a consistent final attribution summary under the empirical-score twin, including the changed T5 isolation verdict.
  2. For the cells currently marked undetermined, what practically meaningful error differences can the available runs exclude at their realized m*, and how sensitive are the claimed equivalences to the margin choice? A precision or minimum-detectable-effect summary would clarify the remaining uncertainty beyond the already reported density sweep and seed-wise matching.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.