When Is Emergent Consensus Real? A Measured Coupling Gain and a Validity Diagnostic for LLM Agent Societies

Abstract: LLM "agent societies" are studied via demonstrations of emergent consensus or polarization -- with no measurable control parameter, no theory of when each regime appears, and no test of whether an outcome is a genuine social dynamic or a model artifact. We introduce the coupling gain gamma, measured per-agent by counterfactually perturbing a neighbour's stated opinion. (i) gamma is stable and model-distinguishing -- across five frontier models it spans 0.15-0.43 (n=20, 95% CIs <= 0.025), paraphrase-invariant; social-neighbour gamma roughly equals numeric-anchor gamma, so gamma is evidence-coupling, not uniquely social. (ii) Classical dynamics with measured (not assumed) coefficients organise the regime: Friedkin-Johnsen for consensus/pluralism, signed-Laplacian/structural-balance for polarization. (iii) Frontier LLMs do not spontaneously backfire (beta <= 0), so default societies do not self-polarize -- polarization is always induced; the beta>0 branch arises only in the FJ surrogate, never in the agents. (iv) A randomized-initial-condition diagnostic -- the (slope, bias) of final vs. initial opinion -- separates genuine averaging from model-prior artifacts (boundary-censoring ruled out by construction via interior-valued facts); applied to a published "emergent consensus" result (Chuang et al. 2023) it reveals a model-specific conflation: averaging on debatable claims, prior-artifact on settled facts. (v) Coupling is context-dependent: pairwise gamma does not predict multi-neighbour outcomes -- it can order them backwards -- whereas a modality-matched group coupling does (sixteen closed+open models, Pearson r=-0.70, permutation p=0.008). The regime laws take this matched coupling, not the single-neighbour gamma: emergent consensus must be read from coupling in the target interaction. We contribute a measurement protocol and a validity instrument, not new theory.

• The paper introduces a novel measurement protocol that quantifies coupling gain (ranging from 0.15 to 0.43) to distinguish genuine emergent phenomena from artifacts in LLM agent societies.
• It applies classical opinion dynamics models to empirically validate consensus, pluralism, and induced polarization through measured coefficients and network simulations.
• A robust diagnostic method separates authentic averaging from model-prior biases, underscoring the non-additive, context-sensitive nature of group influence in LLMs.

Measurement and Validity in Emergent Consensus of LLM Agent Societies

Introduction

This paper, "When Is Emergent Consensus Real? A Measured Coupling Gain and a Validity Diagnostic for LLM Agent Societies" (2606.22203), addresses critical methodological gaps in the study of emergent social dynamics among LLM-based agent societies. The prevailing literature largely focuses on system demonstrations, often lacking control parameters, regime prediction mechanisms, or robust validity checks distinguishing genuine emergent phenomena from artifacts of LLM priors or sycophantic behaviors. To rectify this, the authors introduce an empirical protocol for measuring the agent-level coupling gain and provide formal diagnostics to validate emergent consensus, pluralism, or polarization within these societies.

Coupling Gain: Measurement and Interpretation

The paper introduces the coupling gain $y$, defined as the per-agent sensitivity of opinion updates with respect to a neighbor's stated opinion, measured by counterfactual perturbation. The protocol's robustness is established across five frontier LLMs (DeepSeek, Qwen, Gemini, GPT-5.5, Claude), with $y$ ranging from 0.15 to 0.43 and bootstrap CIs below 0.025. Importantly, $y$ is model-distinguishing, stable under prompt paraphrase, and invariant whether the neighbor's input is social or a numeric anchor. Thus, $y$ is interpreted as an evidence-coupling coefficient—not exclusively a social parameter, but quantifying the agent's susceptibility to external evidence. This distinction corrects prior assumptions conflating sycophantic or homogenizing tendencies with social coupling.

Classical Regime Organization: Pluralism, Consensus, Polarization

Leveraging classical opinion dynamics frameworks (Friedkin-Johnsen, DeGroot, signed-Laplacian for structural balance), the authors establish measured (not assumed) coefficients that organize macro-regimes:

• Consensus vs. pluralism: For FJ dynamics, stationary profiles interpolate between initial opinions and consensus as $y$ increases.
• Polarization: Structural balance theory is applied, with polarization thresholds empirically tested via a signed-Laplacian analysis. Notably, the active-polarization branch ($B > 0$) is not observed in the tested LLM agent societies, contradicting some prior claims—polarization, when observed, is induced and never spontaneous.

These results are formalized and tested via network structures (ring, complete graphs), with stubborn and yielding models behaving in accordance with theoretical predictions: higher $y$ correlates with consensus, lower $y$ preserves pluralism.

Validity Diagnostic: Authenticity vs. Artifact

A novel authenticity diagnostic (slope, bias analysis of final opinion vs. initial opinion) is provided to separate genuine averaging effects from model-prior artifacts. The protocol is censoring-immune via interior-valued numeric facts, ruling out spurious convergence due to boundary effects. Application to published results (e.g., "Simulating Opinion Dynamics with Networks of LLM-based Agents" (Chuang et al., 2023)) exposes confounding mechanisms: emergent consensus on debated claims corresponds to authentic averaging, while settled facts exhibit consensus driven by prior artifacts, with model-specific variation.

Numerical results show bold prior-pull bias for settled facts (e.g., flat-earth, vaccine claims) but neutral/slight bias for debated issues, establishing that the diagnostic cleanly distinguishes artifact-driven convergence from real social dynamics.

Non-Additive Aggregation and Context Dependence

An important finding is the context dependence and non-transferability of coupling gain. While the measurement protocol itself (counterfactual perturbation) transfers across modalities and aggregation schemes, the value of pairwise $y$ does not reliably predict society-level outcomes when agents face multiple neighbors or non-numeric evidence.

• Negative correlation: The susceptibility to coherent group influence (modality-matched group coupling, $p_{\text{ft}}$) can even anti-correlate with pairwise $y$0 (Spearman $y$1 across 15 models), contradicting the predictions of additively separable models (DeGroot, FJ).
• Macro outcome prediction: Only modality-matched group coupling quantitatively predicts society convergence (Pearson $y$2, $y$3 across 16 models), in contrast with pairwise $y$4, which can order societies backwards.
• Aggregation law: The empirical rejection of additivity (Prop. 4) suggests that LLM agent societies exhibit consensus-conditional aggregation, rather than simple linear opinion updating.

These results further imply that emergent consensus must be quantified from coupling measured within the target interaction context, not nominal one-on-one influenceability.

Implications and Future Directions

The paper's measurement protocol and diagnostic instrument provide objective, falsifiable tools for dissecting emergent phenomena in LLM agent societies. Theoretical implications include the necessity to revisit classical opinion dynamics models and aggregation laws to accommodate non-additive, context-sensitive behaviors in high-capacity LLM agents. Practically, these tools can guide the development and benchmarking of LLM-based social simulations, ensuring genuine emergent dynamics are distinguished from model artifacts.

Future developments should expand exploration to larger agent societies, more diverse tasks (beyond opinion-update), and learned network structures, as current limitations stem from small $y$5 (societies of 6–10 agents) and short simulation horizons. Further, closing the mild circularity in prior calibration and robustly quantifying polarization thresholds across architectures remain ripe for extension.

Conclusion

This work establishes a measurement-first paradigm for studying emergent consensus, pluralism, and polarization in LLM agent societies. The coupling gain $y$6 is a stable, model-distinguishing, evidence-coupling coefficient, not uniquely social. Classical regime organization is empirically validated, with induced polarization and non-transferability of pairwise coupling highlighted. The (slope, bias) authenticity diagnostic is censoring-immune and robustly distinguishes real emergent effects from prior artifacts. Society-level outcomes are governed by modality-matched group coupling, not nominal pairwise influenceability, underscoring the context dependence and the inadequacy of additive aggregation models. The measurement protocol and validity instrument introduced are crucial foundations for rigorous evaluation and future theoretical developments in AI agent society research.