Paper deep dive

Belief Coevolution in a Social Network of Generalist and Specialist Large Language Models

Germans Savcisens, Samantha Dies, Courtney Maynard, Tina Eliassi-Rad

Year: 2026Venue: arXiv preprintArea: cs.CLType: PreprintEmbeddings: 123

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%

Last extracted: 8/1/2026, 1:04:35 AM

Summary

This paper introduces CoevolveSim, a framework for studying belief diffusion in multi-agent LLM systems. It isolates the effects of domain specialization, social-role assignment, and network structure on belief formation. Results indicate that LLM heterogeneity (specialists vs. generalists) is the primary driver of collective belief change and consensus shifts, whereas persona-style role assignment and network structure have minimal impact on population-level consensus but reshape individual belief trajectories.

Entities (13)

CoevolveSim · framework · 98%Generalist LLM · agent-type · 95%Llama-3.1-8B-Instruct · model · 95%Llama3-Med42 · model · 95%Specialist LLM · agent-type · 95%Domain Specialization · factor · 92%Network Structure · factor · 92%Social Role · factor · 92%Consensus · metric · 90%Erdős–Rényi Network · network-topology · 90%Watts–Strogatz Network · network-topology · 90%Directedness · metric · 88%plasticity · metric · 88%

Relation Signals (9)

Specialist LLM → includes → Llama3-Med42

confidence 95% · Llama3-Med42 ... is finetuned on clinical text

Generalist LLM → ispoweredby → Llama-3.1-8B-Instruct

confidence 95% · Llama-3.1-8B-Instruct serves as the generalist

CoevolveSim → studies → Belief Diffusion

confidence 95% · We introduce CoevolveSim, a framework for studying belief diffusion within networked LLM populations.

Domain Specialization → drives → Collective Belief Change

confidence 92% · LLM heterogeneity is the primary driver of collective belief change.

Specialist LLM → increases → Consensus Shift

confidence 90% · introducing (finetuned) specialist LLMs more than doubles the shift in consensus

Network Structure → reshapes → Individual Belief Revision

confidence 90% · persona-style role assignment and network structure reshape individual belief revision

Social Role Assignment → reshapes → Individual Belief Revision

confidence 90% · persona-style role assignment and network structure reshape individual belief revision

CoevolveSim → uses → Erdős–Rényi Network

confidence 85% · We represent social connections between agents via two network structures: (1) Erdős and Rényi (1959) (ER)

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Large language models (LLMs) are increasingly deployed in multi-agent environments. However, the processes by which beliefs form and propagate among interacting LLMs remain poorly understood. We introduce CoevolveSim, a framework for studying belief diffusion within networked LLM populations. CoevolveSim allows us to isolate and study three factors: domain specialization, social-role assignment, and social network structure. Within this framework, generalist and specialist LLM agents exchange and revise beliefs. In each round, an LLM agent observes a summary of its neighbors' beliefs before updating its own. We run 1,280 controlled simulations spanning four scenarios, two network structures, and 20 medical-indication statements. We find that persona-style role assignment and network structure reshape individual belief revision but have minimal effect on population-level consensus. In contrast, introducing (finetuned) specialist LLMs more than doubles the shift in consensus and gives rise to consistent asymmetries in exerted influence. We further show that simple persistence-based opinion-dynamics models reproduce collective outcomes in all-generalist LLM populations, whereas heterogeneous LLM populations require population-level belief composition to reproduce consensus and agent identity to predict individual belief transitions. Our results indicate that realistic simulation of belief diffusion in multi-agent LLM systems requires a diverse set of underlying LLMs, not persona prompting alone.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

122,131 characters extracted from source content.

Expand or collapse full text

Belief Coevolution in a Social Network of Generalist and Specialist Large Language Models Germans Savcisens1,2, Samantha Dies1, Courtney Maynard1, Tina Eliassi-Rad1,2,3 1Khoury College of Computer Sciences, Northeastern University, Boston, USA 2Network Science Institute, Northeastern University, Boston, USA 3Santa Fe Institute, Santa Fe, USA Correspondence: g.savcisens@northeastern.edu Abstract Large language models (LLMs) are increasingly deployed in multi-agent environments. However, the processes by which beliefs form and propagate among interacting LLMs remain poorly understood. We introduce CoevolveSim, a framework for studying belief diffusion within networked LLM populations. CoevolveSim allows us to isolate and study three factors: domain specialization, social-role assignment, and social network structure. Within this framework, generalist and specialist LLM agents exchange and revise beliefs. In each round, an LLM agent observes a summary of its neighbors’ beliefs before updating its own. We run 1,2801,280 controlled simulations spanning four scenarios, two network structures, and 20 medical-indication statements. We find that persona-style role assignment and network structure reshape individual belief revision but have minimal effect on population-level consensus. In contrast, introducing (finetuned) specialist LLMs more than doubles the shift in consensus and gives rise to consistent asymmetries in exerted influence. We further show that simple persistence-based opinion-dynamics models reproduce collective outcomes in all-generalist LLM populations, whereas heterogeneous LLM populations require population-level belief composition to reproduce consensus and agent identity to predict individual belief transitions. Our results indicate that realistic simulation of belief diffusion in multi-agent LLM systems requires a diverse set of underlying LLMs, not persona prompting alone. Code and simulation data: carlomarxdk/coevolve-sim. Belief Coevolution in a Social Network of Generalist and Specialist Large Language Models Germans Savcisens1,2, Samantha Dies1, Courtney Maynard1, Tina Eliassi-Rad1,2,3 1Khoury College of Computer Sciences, Northeastern University, Boston, USA 2Network Science Institute, Northeastern University, Boston, USA 3Santa Fe Institute, Santa Fe, USA Correspondence: g.savcisens@northeastern.edu 1 Introduction Figure 1: Schematic overview of CoevolveSim belief diffusion. Given a discussion topic, CoevolveSim first asks each agent for its initial belief without providing information about the neighbors (Round 0). Then, CoevolveSim facilitates the exchange of beliefs and social roles between agents and their immediate neighbors. In each subsequent round, CoevolveSim evaluates the belief of each agent after providing a summary of the beliefs of its neighbors from the previous round. After T rounds of interaction, agents reach their final belief states. In this illustration and in our experiments, we set T=10T=10. Side note: The discussion statement is factually correct. Large language models (LLMs) are becoming integral to the global information ecosystem. Nearly two-thirds of companies now use generative artificial intelligence (AI) to produce written content McKinsey & Company (2025), with AI-generated text appearing in contexts ranging from UN press releases to scientific articles Liang et al. (2025a, b). On social knowledge platforms such as Quora, the share of AI-generated posts increased from 22% in 2022 to 3939% by late 2024 Sun et al. (2025). At the same time, LLMs are increasingly common in multi-agent settings, including debate and collaborative reasoning Guo et al. (2024); Wu et al. (2024); Tran et al. (2025). As these applications grow, understanding how beliefs form and spread within populations of interacting LLMs becomes increasingly important. Questions of collective belief formation have long been studied in the human opinion-dynamics literature. Individuals revise beliefs after exposure to others’ opinions Lorenz et al. (2011) and often reach consensus after repeated interactions Becker et al. (2017). Nevertheless, the degree to which individuals influence others’ beliefs varies unevenly within a population Moussaïd et al. (2013). For example, Bailey et al. (2023) show that a social label assigned to an advisor (whether the advisor is described as an expert or a novice) is the strongest predictor of the extent to which individuals rely on the given advice. Hence, collective outcomes emerge from an interplay between individual characteristics and social structure rather than from any single mechanism in isolation. Existing work on multi-agent LLM systems has largely focused on behavioral outcomes such as task completion, votes, and debate performance Guo et al. (2024). However, LLM-based agents may exhibit varied behaviors as they are powered by different LLMs, assigned certain social roles (e.g., persona-style prompting) or connected to distinct sets of agents in the interaction networks. The relationship between these factors remains largely understudied, making it difficult to quantify their contributions to agent-level beliefs, influence, and population-level dynamics Ng et al. (2026); Ye et al. (2026). We introduce CoevolveSim, a network-mediated belief diffusion framework in which generalist and specialist LLM agents exchange and revise beliefs over repeated rounds of interaction (see Figure˜1). In CoevolveSim, we focus on three factors: 1. Domain specialization, which specifies the LLM that powers an agent and determines its specialization (such as a chemist, a physician, an engineer, etc). 2. Social role, which is assigned to each LLM agent. This persona-style label is communicated to the agent and its neighbors. 3. Structure of the social network, which determines which LLM agents can directly exchange beliefs with each other. This setup allows us to answer two questions, which we study across four experimental scenarios (§4.4): 1. Which factors drive belief diffusion in populations of interacting LLM agents? 2. Can the resulting dynamics be explained using classical opinion-dynamics mechanisms? We quantify the resulting belief dynamics in terms of belief revision, exerted influence, and consensus (§4.5.1). We then assess whether these dynamics can be reproduced by opinion-dynamics models (referred to here as surrogate models) that predict belief transitions using different information, from individual persistence to local social influence and identity-based characteristics (§4.5.2). Our findings: LLM heterogeneity is the primary driver of collective belief change. That is, populations containing a diverse set of LLMs exhibit larger shifts in consensus and stronger influence asymmetries than populations that differ only in social-role assignment (all-generalist populations). Similarly, the factors required to predict individual belief revisions are not necessarily the same as those required to reproduce collective outcomes, highlighting the gap between agent- and population-level explanations of belief diffusion. Contributions 1. We introduce CoevolveSim, a framework for studying belief diffusion in populations of networked LLM agents. 2. We disentangle the effects of domain specialization, social-role assignment, and social network structure through 1,2801,280 controlled simulations across four scenarios. 3. We demonstrate that social-role assignment and network structure each reshape individual belief trajectories, yet leave population-level consensus nearly unchanged (§5.1). 4. We show that LLM heterogeneity is the primary driver of collective belief change, giving rise to opinion leaders (§5.1). 5. We show that the mechanisms that predict individual belief revisions differ from those that reproduce collective consensus (§5.2). 2 Related Work Our work stands at the intersection of (1) research on multi-agent LLM systems, (2) heterogeneous agent populations, and (3) opinion dynamics. 2.1 Beliefs in Multi-Agent LLM Systems Multi-agent LLM systems are typically evaluated through behavioral outcomes such as task completion, voting, and debate performance Guo et al. (2024); Wu et al. (2024); Chen et al. (2024). More recent work focuses directly on the belief states, showing that belief consistency varies across tasks and models Pal et al. (2025). Similarly, population-level outcomes such as consensus can also amplify persuasion Bilgin et al. (2025), further highlighting that interactions between multiple factors affect collective behavior Brockers et al. (2025). However, agent states, local interactions, and population-level dynamics remain jointly understudied Ng et al. (2026), and existing work rarely isolates the effects of domain specialization, social roles, and network structure. 2.2 Domain Specialization and Social Roles Heterogeneity in multi-agent LLM populations is commonly introduced through either prompt-based personas (that instruct an LLM on how to behave) or specialized models (that use various pretrained or finetuned LLMs). Persona-based approaches simulate diverse social identities Chuang et al. (2024); Taubenfeld et al. (2024), but their effects often weaken over extended interactions Bhandari et al. (2025); Villa et al. (2024). Specialized models, by contrast, exhibit greater robustness on domain-specific tasks Tanwar et al. (2025); Chen et al. (2023); Mullick et al. (2024), while homogeneous LLM populations can be vulnerable to consensus collapse and bias amplification Breum et al. (2024); Estornell and Liu (2024); Xiong et al. (2023). So far, existing studies examine personas or specialized models in isolation, making it difficult to separate the effects of portrayed and perceived social roles from those of underlying domain specialization. 2.3 Opinion Dynamics and Network Structure Classical opinion-dynamics models such as those of DeGroot (1974) and Friedkin and Johnsen (1990) describe belief revision as a process of repeated social influence. Empirical studies suggest that local interactions can increase consensus Lorenz et al. (2011); Becker et al. (2017), while members of majority groups or those with specialized knowledge produce asymmetric influence Moussaïd et al. (2013). Network structure further shapes collective outcomes by influencing how information propagates through a population Bakshy et al. (2011). However, network structure is rarely manipulated in multi-agent LLM systems Chuang et al. (2024); Gao et al. (2023). These findings suggest that collective belief dynamics emerge from interactions among specialization, social position, and network structure. Our experimental design operationalizes these three factors within an LLM framework, allowing us to study their individual and combined effects on belief diffusion. 3 Belief Diffusion Framework We introduce CoevolveSim, a network-mediated belief diffusion framework (Figure˜1). We designed it to isolate the effects of three factors: (1) agents’ specialization, (2) social-role assignment, and (3) social network structure. In our framework, generalist LLM agents (powered by LLMs without finetuning) and specialist LLM agents (powered by LLMs finetuned in various fields such as clinical medicine, chemistry, cybersecurity, mathematics, and other fields; see Table˜A2) are embedded in an undirected social network G. In each round t, LLM agent aia_i receives a template message shown in Figure˜2. The message includes the agent’s assigned social role, a discussion statement, and a summary of its neighbors’ beliefs from round t−1t-1. Belief updates are synchronous and Markovian: all agents condition on information from the same previous round, which includes only the latest neighborhood summary. This design allows us to examine how domain specialization, social role, and neighborhood (social network structure) shape beliefs over T=10T=10 rounds. In this section, we define domain specialization and social roles (§3.1), specify network structures (§3.2), formalize belief extraction (§3.3), and define the neighborhood summary (§3.4). Table˜A1 lists the notation used in this section. 3.1 Domain Specialization and Social Roles Each LLM agent is characterized by a domain specialization and a social role. Domain specialization reflects an LLM’s finetuning domain. For example, Llama3-Med42 Christophe et al. (2024) is finetuned on clinical text, so we assume it has a medical specialization. Social role is a social label assigned to an LLM agent and communicated to both the agent and its immediate neighbors. Social roles function as social identities, such as “clinical physician,” “mathematician,” or “chemist.” In our experiments, social roles do not necessarily correspond to specialization. An LLM agent could specialize in medicine but portray itself as a mathematician. This separation allows us to test whether belief dynamics are driven by what an LLM agent knows (domain specialization) or by what it is labeled as (social role). 3.2 Social Networks We embed n LLM agents in an undirected network G, where each agent aia_i is a node and edges represent social connections. Every LLM agent aja_j with an edge to agent aia_i is part of aia_i’s neighborhood (i)N(i). We describe the selection of network structures in §4.3. 3.3 Belief Extraction An LLM agent ℳM maps an input sequence x to a distribution ℳP_M. For any token τ∈ℳτ _M, ℳ(τ∣)P_M(τ x) gives the (uncalibrated) probability that τ is the next token after x. We use ℳP_M together with a templated message x to extract the belief scores Savcisens and Eliassi-Rad (2025). Templated Message. Figure 2: Message structure for belief probing. Each LLM agent receives a message that follows the Hugging Face chat template: (1) a system prompt assigning the agent’s role r x_r and task instructions; (2) a user prompt containing the statement s x_s, the neighborhood summary n(t) x_n^(t) (n0=∅ x_n^0= ), and the multiple-choice question; and (3) the assistant prompt from which we extract the next-token probabilities to obtain belief scores. At round t, LLM agent aia_i receives a message i(t)=(r,s,n(t))m_i^(t)=( x_r, x_s, x_n^(t)) containing the agent’s role r x_r, the discussion statement s x_s, and a summary n(t) x_n^(t) of neighbors’ beliefs from round t−1t-1, with n(0)=∅ x_n^(0)= . The message ends with a multiple-choice question asking the agent to classify the statement s x_s as (1) correct, (2) incorrect, or (3) neither (see Figure˜2). Belief. Given message i(t)m_i^(t), we operationalize LLM agent i’s belief at round t as i(t)(y∣i(t))=ℳ([ky]∣i(t)), b_i^(t)\! (y _i^(t) )=P_M\! ( [k_y ] _i^(t) ), (1) where ky∈1,2,3k_y∈\1,2,3\ indexes y∈y∈\correct, incorrect, neither\. The next-token probability assigned to the corresponding option is the belief score. Additionally, any residual probability mass not captured by these tokens is added to i(t)(neither) b_i^(t)(neither) so that the three scores sum to 11. Further, the LLM agent’s discretized belief is Bi(t)=arg⁡maxy⁡i(t)(y).B_i^(t)= _y b_i^(t)(y). (2) 3.4 Neighborhood Summary At round t, we collect the previous round’s discretized beliefs Bj(t−1)B_j^(t-1) from aia_i’s neighbors and group the neighbors into three categories: agree, disagree, and unsure. For each group, we generate a sentence listing the corresponding neighbor roles and counts (see Figure˜2), and concatenate the three sentences to form a summary of the neighbors’ beliefs n(t) x_n^(t). To quantify the initial belief of each LLM agent, no neighborhood summary is provided at t=0t=0. To reduce the ordering bias, we randomly shuffle social roles within each sentence and randomize sentence ordering before inserting them into the templated message. 4 Experimental Setup In this section, we describe the LLMs used in our study (§4.1), discussion statements (§4.2), network generation (§4.3), simulation scenarios (§4.4), as well as details on the descriptive evaluation (§4.5.1) and surrogate analysis (§4.5.2). 4.1 Large Language Models To study how domain specialization is associated with belief formation, we use 1515 LLMs from the Llama-3 Grattafiori et al. (2024) family. Llama-3.1-8B-Instruct serves as the generalist, since it is pretrained without a domain-specific corpus. The other 1414 LLMs serve as specialists and are finetuned on domain-specific corpora, such as biomedical texts, cybersecurity reports, and programming tasks. Each LLM has 88 billion parameters, so performance differences are more likely to come from finetuning rather than model size. Table˜A2 provides information about the LLMs used in our experiments. 4.2 Discussion Statements We sample 2020 affirmative statements referencing real-world medical indications from the corpus introduced by Savcisens and Eliassi-Rad (2025). An example of a sampled true statement is “Terbutaline is indicated for the treatment of asthma,” while “Ethinylestradiol is indicated for the treatment of dry cough” is an example of a false statement. We use zero-shot predictions from a medical LLM111Llama3-Med42 Christophe et al. (2024) is hereafter referred to as the medical LLM. and 1414 additional LLMs to construct features that capture (1) the binary ground-truth label of a statement (true, false), (2) label predicted by the medical LLM, (3) accuracy of the medical LLM, (4) average accuracy of other LLMs, and (5) agreement between the medical LLM and other LLMs. We apply the maximin criterion Johnson et al. (1990) to sample a diverse set of statements. That is, we maximize the minimum pairwise distance in this feature space. We provide additional details and a list of statements in Appendix˜C and Table˜A4. 4.3 Social Networks We represent social connections between agents via two network structures: (1) Erdős and Rényi (1959) (ER), and (2) Watts and Strogatz (1998) (WS). In ER, G(n,p)G(n,p), each pair of LLM agents is connected with a probability p. It typically produces locally tree-like networks with low clustering and light-tailed degree distributions Barabási and Pósfai (2016). In WS, G(n,β,k)G(n,β,k), the network is constructed by rewiring edges of a k-regular ring lattice with probability β. It generally produces networks with high local clustering and short average path lengths. We fix n=48n=48 for both network structures, with p=0.3p=0.3 for ER Leskovec (2017) and k=8k=8, β=0.1β=0.1 for WS. We select 88 network instances using the maximin criterion described in Section˜C.2 for each network structure and refer to these as network realizations.222In total, we have 1616 network realizations. 4.4 Scenarios and Simulations We examine belief diffusion across four scenarios. These scenarios isolate the effects of domain specialization, social-role assignment, and role–specialization alignment. I. Baseline Generalists. All LLM agents use the same generalist LLM and receive the same social role, serving as a no-specialization, no-role baseline. I. Generalists with Random Roles. All LLM agents use the same generalist LLM, but each is assigned a random social role drawn from Table˜A2. This isolates the effect of the role assignment. I. Specialists with Random Roles. LLM agents are powered by generalist and specialist LLMs with randomly assigned roles. The list of LLMs and social roles is provided in Table˜A2. This scenario isolates the effect of heterogeneity in specialization. IV. Specialists with Matched Roles. LLM agents are powered by generalist and specialist LLMs. Further, we assign social roles that match LLMs’ capabilities (e.g., an agent powered by the math LLM is assigned the role of “Mathematician”). Compared with scenario I, here we isolate the marginal effect of role–specialization alignment. Runs. Each scenario is simulated multiple times; we refer to each individual simulation as a run. Each run begins with an initial round t=0t=0, where we collect LLM agents’ initial beliefs before introducing neighbor information. It is followed by T=10T=10 synchronous belief-update rounds, yielding 11 states t=0,…,Tt=0,…,T. In total, we perform 1,2801,280 runs: 44 scenarios, 22 network structures with 88 realizations each, and 2020 discussion statements. LLMs / Specialization Social Roles I Generalist Only “LLM” I Generalist 1515 Random roles I Generalist & 1414 Specialists 1515 Random roles IV Generalist & 1414 Specialists 1515 Matched roles Table 1: Composition of scenarios. Specialization and Social Roles show the number of unique LLMs and social roles, respectively. Scenarios I vs. I isolate the effect of the persona-style role assignment; I, I vs. I, IV isolate the effect of LLM heterogeneity (introduction of specialists); I vs. IV isolate the effect of role–specialization alignment. These four scenarios allow us to isolate the effects of social role and specialization and are summarized in Table˜1. Comparing scenarios I with I isolates persona-style role assignment; comparing I with I isolates the introduction of specialists (the specialization effect); and comparing I with IV isolates role–specialization alignment. We also report a fourth contrast, the composition effect, for completeness only; it averages over role assignment to contrast the homogeneous scenarios I, I with the heterogeneous I, IV, but our analysis relies on the specialization effect (I→ ), which isolates the introduction of specialists with role assignment being fixed. 4.5 Evaluation Our analysis consists of two stages. We first characterize belief trajectories and compare dynamics across scenarios and network structures (§4.5.1). We then evaluate the extent to which classical opinion-dynamics models explain the observed LLM belief dynamics using a hierarchy of increasingly expressive surrogate models (§4.5.2). 4.5.1 Metrics and Dynamics Analysis We characterize belief trajectories using three agent-level metrics: magnitude of belief change (plasticity), directional consistency of these changes (directedness), and outgoing influence. Additionally, we consider the consensus: a population-level outcome that quantifies the degree of agreement among all LLM agents in the network. Agent-level metrics. We first examine the metrics that quantify agent-level belief trajectories. Hence, we define agent aia_i’s per-round belief change: Δi(t)=i(t)−i(t−1). b_i^(t)= b_i^(t)- b_i^(t-1). (3) Plasticity is the magnitude of an LLM agent’s belief change per round (averaged over a run): Plasticity(i)=1T∑t=1T12‖Δi(t)‖1Plasticity(i)= 1T _t=1^T 12 \| b_i^(t) \|_1 (4) LLM agents with high plasticity revise beliefs frequently or undergo several large-magnitude changes in belief, whereas agents with low plasticity remain stable. Directedness captures whether an LLM agent revises belief in a consistent direction: Directedness(i)=‖i(T)−i(0)‖1∑t=0T−1‖i(t+1)−i(t)‖1Directedness(i)= \| b_i^(T)- b_i^(0) \|_1 _t=0^T-1 \| b_i^(t+1)- b_i^(t) \|_1 (5) Directedness closer to 11 indicates that an agent’s beliefs move toward a single label (e.g., increasing the probability assigned to a correct label) across rounds, while 0 indicates belief oscillation.333The denominator is 0 only for an LLM agent whose belief never changes, yielding an undefined ratio; our implementation is designed to exclude such cases, however, we do not observe this in practice. Plasticity and directedness characterize the shape of individual belief trajectories. An LLM agent that updates frequently but oscillates (high plasticity, low directedness) follows a different trajectory than one that updates frequently and monotonically (high plasticity, high directedness). Outgoing Influence. The agent-level metrics characterize belief trajectories in isolation and cannot establish whether some agents influence or follow the beliefs of their neighbors, i.e., the emergence of opinion-leader-like agents (LLM agents whose own belief changes are associated with larger subsequent changes among beliefs of their neighbors). To capture the influence, we measure the average magnitude of neighbors’ belief changes at t+1t+1, weighted by how much the agent aia_i itself moved at t.444Rounds in which aia_i has a small change in belief make a smaller contribution to the outgoing influence score. A higher influence in Equation˜7 indicates that belief updates in aia_i are associated with greater subsequent updates among its neighbors (i)N(i). Let δi(t)=12‖Δi(t)‖1 _i^(t)= 12 \| b_i^(t) \|_1 (6) denote aia_i’s absolute per-round belief change. The outgoing influence is then defined as Influence(i)=∑t=0T−1[δi(t)⋅∑j∈(i)δj(t+1)]|(i)|⋅∑t=0T−1δi(t).Influence(i)= _t=0^T-1 [ _i^(t)\,·\, _j (i) _j^(t+1) ]|N(i)|\,·\, _t=0^T-1 _i^(t). (7) To assess whether particular agents exert more influence than others, we fit a variance-components model Scheipl et al. (2008) to the outgoing influence. By treating LLM agents, network realizations, and statements as random effects, we can quantify how much variance in the outgoing influence is attributable to each. For each random-effects component, we report the intraclass correlation coefficient (ICC), which is the proportion of total variance attributable to that component. A higher agent-level ICC suggests the presence of opinion leaders555Opinion-leader-like agents are LLM agents with persistently high outgoing influence. and followers. We provide full model specifications in Appendix˜D. Population-level metrics. We track consensus to estimate the population-level agreement among the discretized beliefs of LLM agents: Cons(t)=1n(n−1)∑i≠j[Bi(t)=Bj(t)].Cons(t)= 1n(n-1) _i≠ j 1\! [B_i^(t)=B_j^(t) ]. (8) In our analysis, we focus on the change in consensus between the initial (t=0t=0) and final rounds. A positive value indicates that belief exchange moved the population closer to unanimous agreement, while a negative value indicates that beliefs became more fragmented. Together, these provide a comprehensive description of belief diffusion that no single metric can provide alone. For instance, a run with low change in consensus may reflect either near-zero change in beliefs (low plasticity) or frequent oscillation without convergence toward agreement (high plasticity with low directedness). Further, if plasticity and directedness both have similar values across all LLM agents, the outgoing influence shows whether some LLM agents’ belief updates are persistently associated with larger subsequent updates among their neighbors (opinion-leader-like agents). Statistical Inference. To account for repeated measures across statements and network realizations, we fit linear mixed-effects models. Here, we treat scenarios (I–IV) and network structures (ER and WS) as fixed effects, while statements and network realizations are treated as random effects. We apply this design to plasticity, directedness, and change in consensus, and we report estimated marginal means (EMMs) with 9595% confidence intervals in §5. Note that agent-level metrics are averaged within each run (n=1,280n=1,280). Contrast Analysis. To quantify the effect sizes between scenarios, we perform a contrast analysis. The contrasts include the role effect (I vs. I), specialization effect (I vs. I), and role–specialization alignment (IV vs. I); these are summarized in Table˜A5. For each contrast, we report three quantities: 1. the marginal contrast (Table˜A8), which averages effects across network structures; 2. the conditional contrasts (Table˜A9), which estimate the same effect separately within ER and WS networks; 3. the interaction contrast (Between rows in Table˜A9), which tests whether a manipulation produces a different effect across two network structures: a non-significant interaction indicates that we do not detect any differences in the effects. All contrasts are reported as Cohen’s d with 95%95\% confidence intervals. We provide full model specifications and contrast definitions in Appendix D. 4.5.2 Surrogate Models of Belief Diffusion To evaluate whether the observed belief dynamics can be explained by classical opinion-dynamics mechanisms, we fit a hierarchy of surrogate transition models. Each surrogate predicts an LLM agent’s discretized belief at round t+1t+1 from information available at round t. The hierarchy progressively introduces mechanisms such as persistence, social belief composition, and the agent’s identity, allowing us to test which mechanisms are necessary to reproduce the observed dynamics. Formally, each surrogate estimates Pr⁡(Bi(t+1)=y∣ϕi(t)), \! (B_i^(t+1)=y φ_i^(t) ), (9) where Bi(t)∈correct,incorrect,neitherB_i^(t)∈\ correct, incorrect, neither\ denotes the discretized belief of agent aia_i at round t, and ϕi(t) φ_i^(t) is a feature vector encoding information available prior to the update. Model Features ϕi(t) _i^(t) Mechanism M1 Bi(t)B_i^(t) Individual persistence M2 Bi(t),Γ(t)B_i^(t), ^(t) global belief composition M3 Bi(t),Ni(t)B_i^(t),N_i^(t) Local belief composition M4 Bi(t),Ni(t),Mi,RiB_i^(t),N_i^(t),M_i,R_i Local influence + identity Table 2: Hierarchy of surrogate models. Γ(t) ^(t) denotes the population-level belief distribution at round t, Ni(t)N_i^(t) denotes the belief distribution within agent i’s neighborhood, MiM_i denotes the agent’s underlying domain specialization, and RiR_i denotes its assigned social role. M1 serves as an empirical baseline. It captures the extent to which belief updates can be explained by an agent’s current beliefs alone Bi(t)B_i^(t), without access to any other information. M2 augments the baseline with a population-level distribution over belief states Γ(t) ^(t). Each component of Γ(t) ^(t) captures the proportion of LLM agents expressing discretized beliefs at round t. Therefore, M2 is closely related to the population-averaging model of DeGroot (1974). M3 instead incorporates the (local) neighborhood-level distribution over belief states Ni(t)N_i^(t). Each component of Ni(t)N_i^(t) captures the proportion of aia_i’s neighbors expressing discretized beliefs at round t. This surrogate is analogous to opinion-dynamics models with localized social belief composition Friedkin and Johnsen (1990). Finally, M4 extends M3 by incorporating an agent’s social role RiR_i and underlying LLM MiM_i. It captures the effect of the LLM agent’s identity on the transition dynamics. The hierarchy of surrogate models in Table˜2 serves as a controlled ablation. It allows us to identify which of the mechanisms (persistence, social belief composition, or agent identity) are necessary to replicate the observed belief dynamics. Evaluation of surrogates. We evaluate surrogate models using two complementary criteria. First, we assess one-step predictive performance using the Matthews Correlation Coefficient (MCC), a balanced classification metric that accounts for all entries of the confusion matrix Chicco and Jurman (2020). Higher MCC values indicate a more accurate prediction of individual belief updates. Second, we evaluate whether a surrogate reproduces population-level dynamics. We fit each surrogate on empirical transitions, simulate belief diffusion using the fitted model, and compare the resulting consensus trajectory to the original simulation. We quantify this agreement using consensus fidelity, CF=1−1R∑r=1R|Cons^r(T)−Consr(T)|,CF=1- 1R _r=1^R | Cons_r(T)-Cons_r(T) |, (10) where Consr(T)Cons_r(T) and Cons^r(T) Cons_r(T) denote the empirical and surrogate consensus levels at the final round T of run r, respectively. Higher consensus fidelity indicates that the surrogate more accurately reproduces the collective outcome. Together, MCC values and consensus fidelity distinguish between two forms of success: accurately predicting individual belief updates and accurately reproducing population-level belief dynamics. Additional implementation details, feature definitions, fitting procedures, and rollout protocols are provided in Appendix E. 5 Results We first examine which factors drive belief diffusion in LLM populations by comparing the effects of domain specialization, social role assignment, and network structure on belief updating, influence, and consensus formation (§5.1).666When comparing scenarios or network structures, we also report a standardized effect size via Cohen’s d, computed from contrasts. Table A8 shows marginal effect sizes, and Table A9 provides conditional and interaction effect sizes. We then ask whether the resulting belief dynamics can be explained by a hierarchy of classical opinion-dynamics surrogate models (§5.2). 5.1 What drives belief revisions? We look at two levels of belief dynamics: the agent-level (how individual LLM agents revise their beliefs) and the population-level (consensus). Across four scenarios, these two levels dissociate. Social role assignment (scenario I) reshapes individual belief trajectories but produces only a modest effect on population-level consensus (§5.1.1); LLM heterogeneity (scenarios I and IV) produces the largest changes in population-level consensus (§5.1.2–§5.1.3), while aligning role labels to the underlying specialization (IV) adds little at either level (§5.1.4); finally, network structure modulates agent-level behavior while having a negligible effect on the population-level outcome (§5.1.5). We provide detailed results in Tables˜A6, A8, A9 and A7. Figure 3: Belief dynamics across settings and network structures. Estimated marginal means (with 95%95\% confidence intervals) from mixed-effects models. Spcs stands for specialists. (A) Plasticity: average per-round magnitude of belief change. Increases from baseline (I) through random specialists (I); role alignment (IV) reduces plasticity in ER but increases it in WS. (B) Directedness: consistency of belief updates across rounds. Drops sharply once roles are introduced (I→ ) and reaches its minimum in WS with matched specialists (IV). (C) Consensus change: shift in overall belief agreement between the initial and last rounds. LLM heterogeneity drives the largest shifts in consensus. 5.1.1 Social roles change how LLM agents revise beliefs Plasticity (Figure˜3A) is lowest in the baseline scenario (I) with estimated means of 0.030.03 for ER and 0.030.03 for WS. Assigning social roles to generalist agents (I→ ) increases plasticity to 0.070.07 in ER and 0.050.05 in WS, with the marginal role effect d=0.32d=0.32. Scenario I also produces the largest drop in directedness, from 0.970.97 to 0.440.44 in ER and 0.960.96 to 0.500.50 in WS, with the marginal role effect d=−1.86d=-1.86. LLM agents exhibit more belief “movement,” but this movement is not monotonic (Figure˜3B): the few revisions that occur do not move toward a specific belief label. This is further supported by the change in consensus in Figure˜3C, where scenario I produces a minimal change in consensus of 0.040.04 for ER and 0.050.05 for WS, with the marginal effect d=0.51d=0.51. Social roles (persona-style prompting) substantially alter how LLM agents revise beliefs but produce only a modest consensus shift relative to the baseline scenario (I). 5.1.2 LLM heterogeneity is a driver of collective change Introducing specialist LLMs alongside generalists (I→ ) further increases plasticity from 0.070.07 to 0.130.13 in ER and from 0.050.05 to 0.110.11 in WS, with the marginal (I→ ) effect d=0.53d=0.53. Here, directedness falls to 0.270.27 in ER and 0.350.35 in WS with the marginal effect d=−0.62d=-0.62. Likewise, scenario I results in a higher consensus change (Figure˜3C) with 0.190.19 for ER and 0.180.18 for WS, with the marginal (I→ ) effect d=1.39d=1.39. Unlike social role assignment (I), introducing specialist LLMs (I) has a larger effect on both agent- and population-level dynamics. 5.1.3 LLM heterogeneity introduces between-agent differences in influence The shift from scenario I→ is accompanied by the emergence of opinion-leader-like agents (Figure˜4). In generalist scenarios (I, I), the agent-level ICC777Agent-level ICC quantifies the proportion of variance in the outgoing influence attributable to the LLM agent identity. for outgoing influence is statistically indistinguishable from zero, indicating that no LLM agents systematically exert more influence than others. Once specialist LLMs are introduced, the agent-level ICC rises to 4.974.97% in scenario I and 2.732.73% in scenario IV (p<0.001p<0.001 for both). That is, introducing LLM heterogeneity produces a subset of agents with persistent between-agent differences in outgoing influence. We do not observe this asymmetry in persona-style role assignment with generalist LLMs (scenario I). Figure 4: Variance decomposition of outgoing influence. Intraclass correlations (with 95%95\% confidence intervals) from mixed-effects models with random intercepts for agent, network realization, and discussion statement. Spcs stands for specialists. Agent component shows the proportion of total outgoing influence variance attributable to agent-level differences; the network component measures variance attributed to network realizations. Statement and residual components are not shown. Agent component is statistically zero in generalist LLM scenarios (I, I) and rises to 4.97%4.97\% and 2.73%2.73\% in specialist scenarios (I and IV, respectively; Table˜A7). That is, agent-level differences in outgoing influence emerge only when specialist LLMs are present. 5.1.4 Role-specialization alignment leaves trajectories largely intact but dampens the consensus change Aligning social roles with the underlying domain specialization (I→ ) has negligible marginal effects on belief trajectories (Figure˜3): the marginal effect d=0.06d=0.06 on plasticity and d=−0.09d=-0.09 on directedness. However, we observe a small marginal effect in consensus change (d=−0.36d=-0.36), where values drop (I→ ) from 0.190.19 to 0.160.16 in ER and from 0.180.18 to 0.140.14 in WS.888Agents still move about three times farther from their initial agreement than under random roles alone (I→ ). Once LLM heterogeneity is introduced (I→ ), further alignment of social labels with specialization (I→ ) has little effect on agent-level trajectories, reduces rather than increases the consensus (Figure˜3C) and does not amplify the leader–follower effect (Figure˜4). 5.1.5 Social roles interact with the network structure at the agent level but not at the population level The effect of introducing specialist LLMs does not differ between ER and WS networks: the (I→ ) interaction contrast is non-significant for every outcome (plasticity p=0.94p=0.94, directedness p=0.38p=0.38, outgoing influence p=0.26p=0.26, and consensus change p=0.10p=0.10). However, social role assignment (I→ ) is modulated by network structure: the increase in plasticity is twice as large in ER (d=0.43d=0.43) as in WS (d=0.21d=0.21), with an interaction effect d=0.22d=0.22. Similarly, the directedness has a larger drop in ER (d=−1.99d=-1.99) than in WS (d=−1.74d=-1.74), with an interaction effect d=−0.26d=-0.26. Role–specialization alignment (I→ ) is the only manipulation for which network structure produces opposite effects at the agent level: plasticity decreases in ER (d=−0.15d=-0.15) but increases in WS (d=0.26d=0.26), with an interaction effect d=−0.41d=-0.41; while directedness increases in ER (d=0.13d=0.13) but decreases in WS (d=−0.32d=-0.32), with an interaction effect d=0.45d=0.45. Consensus change, by contrast, is network-structure-invariant: the interaction effect is non-significant for every contrast (role, I→ : p=0.21p=0.21; LLM heterogeneity, I→ : p=0.10p=0.10; role–specialization alignment, I→ : p=0.95p=0.95). Network structure, therefore, reshapes how individual agents revise their beliefs but does not affect where the agents collectively converge. Figure 5: Final-state MCC across scenarios and surrogate models. It measures the accuracy of the surrogate’s predicted final belief against the observed final belief. Spcs stands for specialists. Each panel corresponds to one of four experimental settings: (A) Base LLMs, (B) Random roles, (C) Specialists with random roles, and (D) Specialists with matched roles. Each line connects mean estimates (with 95% bootstrap confidence intervals) across surrogate models of increasing complexity for Erdős–Rényi (ER) and Watts–Strogatz (WS). Surrogate models with additional social information (M3 and M4) provide limited benefit in generalist scenarios but modestly improve predictive accuracy in specialist scenarios (I and IV). We provide detailed statistics in Table˜A10. Figure 6: Consensus fidelity across scenarios and surrogate models. Spcs stands for specialists. Each panel corresponds to one of four experimental settings: (A) Base LLMs, (B) Random roles, (C) Specialists with random roles, (D) Specialists with matched roles, and each line connects mean estimates (95% bootstrap confidence intervals) across surrogate models of increasing complexity for Erdős–Rényi (ER) and Watts–Strogatz (WS) networks. In generalist scenarios (I, I), simpler surrogates achieve the highest consensus fidelity, while specialist scenarios (I, IV) require at least global or local belief composition to reproduce observed consensus dynamics. We provide detailed statistics in Table˜A11. 5.2 Can classical opinion-dynamics models explain these dynamics? Section 5.1 established a dissociation between agent- and population-level dynamics: social roles primarily reshape how agents update their beliefs, while LLM heterogeneity additionally produces changes in where the population converges (changes in consensus). We now ask what mechanisms are needed to predict individual belief transitions and to reproduce population-level outcomes (consensus). We evaluate the surrogate hierarchy introduced in §4.5.2, where M1 captures individual persistence; M2 combines individual persistence with the population-level belief composition; M3 combines individual persistence with the local neighborhood belief composition; and M4 extends M3 with an LLM agent’s domain specialization and social-role information. We provide an overview of the results in Tables˜A10 and A11. 5.2.1 Generalist populations are explained by simple dynamics The generalist LLM scenarios (I, I) exhibit high predictive performance across the entire surrogate hierarchy. In scenario I (Figure˜5A), the final-state MCC is high under M1 with 92.4%92.4\% for ER and 90.290.2% for WS; and in scenario I, M1 achieves 86.1%86.1\% for ER and 94.7%94.7\% for WS (Figure˜5B). Overall, the performance gain from M1→ 4 is modest in both scenarios (I: +4.8+4.8 percentage points (p) for ER, +2.3+2.3 p for WS; I: +4.2+4.2 p for ER and 0.00.0 p for WS). Consensus fidelity reveals that individual persistence (M1) is sufficient to explain population-level dynamics in generalist scenarios (Figure˜6A–B). In the baseline scenario (I), M1 achieves the highest consensus fidelity (91.1%91.1\% ER and 90.4%90.4\% WS), with performance decreasing once additional features are introduced (M4: 81.0%81.0\% for ER and 78.5%78.5\% for WS). In the random-role scenario (I), M2 performs best (84.7%84.7\% for ER and 86.7%86.7\% for WS), but its confidence intervals overlap with M1 (80.1%80.1\% ER, 82.1%82.1\% WS; see Table˜A11). Here, surrogates with richer features (M3, M4) also have lower consensus fidelity, e.g., M4 achieves 77.6%77.6\% ER and 76.6%76.6\% WS (Figure˜6B). Individual persistence and the population’s belief composition are sufficient to capture agent-level belief transitions in generalist scenarios. Explicitly modeling local neighborhood composition or agent identity provides little additional explanatory power. Meanwhile, consensus in scenarios I and I appears to emerge from nearly independent belief trajectories: adding social features (M2–M4) does not improve consensus fidelity. This is consistent with the descriptive findings in §5.1.1. 5.2.2 Heterogeneous LLM populations require belief composition Descriptive evaluation (§5.1.2) showed that specialist scenarios (I, IV) produced richer belief dynamics than the generalist scenarios (I, I). We now ask whether these dynamics require more expressive surrogate models. Final-state MCC increases with feature richness (Figure˜5C–D): in scenario I from 76.3%76.3\% (ER) and 84.4%84.4\% (WS) under M1 to 82.5%82.5\% and 89.1%89.1\% under M4. We observe a similar pattern in scenario IV: from 81.7%81.7\% (ER) and 79.6%79.6\% (WS) to 84.7%84.7\% and 82.0%82.0\%. Hence, domain specialization and social-role features enhance prediction of individual belief transitions. Consensus fidelity, however, exhibits a different pattern. M1 achieves 72.1%72.1\% (ER) and 73.3%73.3\% (WS) in scenario I, and 73.7%73.7\% (ER) and 73.4%73.4\% (WS) in scenario IV. Adding population-level belief composition (M2) increases consensus fidelity to ≥92%≥ 92\% across both scenarios and network structures, while M3 and M4 provide no further improvements (confidence intervals overlap; Table˜A11). Here, surrogate analysis shows that heterogeneous LLMs produce belief dynamics that cannot be captured by persistence alone. Features tied to the collective belief composition and agent identity (domain specialization and social roles) shape how individual agents revise beliefs, and thus improve prediction of individual transitions. This extends the agent- and population-level dissociation in §5.1. However, unlike the generalist scenarios (I, I), where individual persistence (M1) largely reproduces consensus, the heterogeneous scenarios (I, IV) require features that account for the “collective” belief composition: consensus emerges as a process in which belief revisions depend on the beliefs circulating in the population or neighborhood. 5.2.3 Agent-level prediction does not guarantee population-level fidelity The two evaluation metrics recover the agent–population dissociation in §5.1. In heterogeneous scenarios (I, IV), final-state MCC generally increases with feature richness, peaking at M4 (I: 82.5%82.5\% ER, 89.1%89.1\% WS; Figure˜5C–D). However, these gains are not mirrored in consensus fidelity, where M2–M4 perform nearly identically (M2 for I: 94.5%94.5\% ER and 93.3%93.3\% WS; Figure˜6C–D). The same gap appears in homogeneous LLM scenarios (I, I), but in the opposite direction. Consensus fidelity is highest under the simplest surrogate and decreases in more complex surrogates. For example, in scenario I, fidelity decreases from 91.1%→81.0%91.1\%→ 81.0\% (ER; Figure˜6A); however, final-state MCC has a modest increase as we add more features: 92.4%→97.2%92.4\%→ 97.2\% (Figure˜5A). 5.2.4 Network structure informs individual transitions but not consensus Finally, surrogate analysis mirrors §5.1.5: network structure primarily influences how individual agents revise their beliefs while leaving collective outcomes unchanged. Final-state MCC differs between ER and WS networks: in scenario I, WS has consistently higher scores than ER (by 8.18.1, 4.14.1, 5.85.8, and 6.66.6 percentage points across M1–M4; Figure˜5C). Comparable separation appears in scenarios I and I (Figure˜5A–B). Consensus fidelity, by contrast, shows no such separation: the ER and WS point estimates are close and their confidence intervals overlap (see Figure˜6 and Table˜A11) across every scenario (I–IV) and surrogate (M1–M4). 6 Discussion Finding Evidence from §5.1 Evidence from §5.2 Social roles change how LLM agents revise beliefs Introducing random social roles (I→ ) alters agent-level trajectories but has much smaller effects on consensus than introducing specialist LLMs (I→ ) (Figure˜3). Generalist scenarios (I, I) are well explained by simple surrogates, whereas specialist scenarios (I, IV) require collective belief composition to reproduce consensus (Figures˜5 and 6). LLM heterogeneity drives both agent- and population-level dynamics Introducing specialist LLMs (I→ ) changes both agent- (plasticity) and population-level (consensus) dynamics (Figure˜3). Specialist scenarios differ from generalist scenarios in both transition predictability and consensus fidelity (Figures˜5 and 6). Specialist LLMs produce a leader–follower structure Variance in outgoing influence attributed to the agent identity is significant only in specialist scenarios (Figure˜4). Identity and local-neighborhood features improve individual transition prediction (final-state MCC) primarily in specialist scenarios (Figure˜5C–D). Role–specialization alignment leaves trajectories largely intact but dampens the consensus change Matching social roles to specialization (I→ ) has only minor effects on belief dynamics (Figures˜3 and 4). Surrogate model predictability (final-state MCC, Figure˜5C–D) and consensus fidelity (Figure˜6C–D) are similar across the two specialist scenarios. Network structure further modulates agent- but not population-level outcomes Network structure modulates agent-level belief revision under the role (I→ ) and alignment (I→ ) manipulations, but the effect of specialist LLMs (I→ ) is network-invariant (Table˜A9). ER and WS differ in final-state MCC (Figure˜5), but not in consensus fidelity. Table 3: Summary of the core findings. The dynamics analysis (§5.1) identifies factors associated with belief diffusion in LLM populations, while the surrogate analysis (§5.2) evaluates which opinion-dynamics mechanisms are required to reproduce them. Despite addressing different questions, both analyses converge on the same qualitative conclusions. Our results suggest that belief diffusion in networked LLM populations is driven more by underlying LLM heterogeneity than by persona-style prompting alone. We provide a summary of findings in Table˜3. Persona-style role assignment is a weak driver of population-level diversity The predominant way to introduce behavioral diversity in LLM multi-agent simulations is through persona-style role assignment: assigning distinct social roles to otherwise identical LLMs Chuang et al. (2024); Taubenfeld et al. (2024). Our analysis shows that social roles do reshape individual trajectories: LLM agents revise their beliefs more often, but instead of converging monotonically toward a specific belief label, they oscillate. This added movement does not propagate to the population level, as the consensus shift stays far below that of introducing heterogeneous LLMs (Table˜A8). Persona-style prompting therefore alters how individual LLM agents update, but it does not by itself produce substantial population-level diversity: it “excites” individual LLM agents without carrying them far from their baseline beliefs That is, assigning roles is not sufficient to simulate a population of agents with diverse behavior. LLM heterogeneity is the stronger driver of population-level change The effect of introducing specialist LLMs among the generalists (I→ , d=1.39d=1.39) on consensus change is 2.7×2.7× that of a social-role assignment (I→ , d=0.51d=0.51; Table˜A8). Because specialists are finetuned on different data and with different methods, they bring different knowledge priors into the interaction. While persona labels (I) only rename otherwise identical “belief generators”, specialists introduce diverse beliefs. In heterogeneous scenarios (I, IV), LLMs differ not only in their embedded knowledge, but also in how they respond to their neighbors’ beliefs. LLM agents exhibit patterns of outgoing influence that are absent in all-generalist scenarios (I, I): some adopt their neighbors’ beliefs, while others hold to their initial position (emergence of opinion-leader-like agents, Figure˜4). The surrogate analyses explain this distinction (§5.2.2). For homogeneous scenarios (I, I), simpler surrogate models reproduce the observed dynamics, but heterogeneous scenarios (I, IV) require surrogates that incorporate evolving belief composition (not persistence alone) to recover the collective outcome. Network structure primarily modulates individual belief revision Network structure further shapes the belief trajectories, but its effect is more limited than that of LLM heterogeneity. That is, network structure changes how individual LLM agents revise beliefs and how outgoing influence is distributed, yet it does not affect the consensus (Table˜A8). In particular, the evaluation metrics for ER and WS networks separate more clearly in agent-level outcomes than in consensus, indicating that network structure mainly affects who revises beliefs and who influences whom. This pattern is consistent with the surrogate analysis. For individual belief transitions, incorporating network information improves prediction in heterogeneous scenarios (especially when agent identity and local neighborhood composition are included). By contrast, consensus fidelity is insensitive to adding more detailed features, and the main gains come from accounting for population-level belief composition (§5.2.2). Agent-level and population-level explanations are only partially aligned The factors that best predict individual belief transitions are not always the same as those needed to reproduce population-level consensus. These findings align with research in human opinion dynamics, which demonstrates that belief diffusion is influenced by both social factors and the structure of the interacting population Lorenz et al. (2011); Becker et al. (2017). Specifically, human-centered studies have established that influence is often distributed unevenly and that local network interactions can shape collective agreement in ways not captured by simple contagion models Moussaïd et al. (2013); Bakshy et al. (2011). Our results reveal a similar pattern in LLM populations: while role labels and neighborhood structure affect individual updating, the underlying composition of the LLM agent population is the primary driver of changes in collective consensus. Our results suggest that the mechanisms governing individual belief revision and collective belief formation overlap only partially, making both levels of analysis necessary for understanding belief diffusion in LLM populations. 7 Conclusion CoevolveSim allows us to disentangle the effects of domain specialization, social-role assignment, and network structure on belief diffusion within populations of interacting LLM agents, and answer the following two questions: 1. Which factors drive belief diffusion? LLM heterogeneity, rather than persona-style role assignment, emerges as the primary driver of collective change. Introducing specialist LLMs more than doubles the shift in consensus compared to social role labels alone and leads to the emergence of high-influence LLM agents, a pattern not observed in all-generalist scenarios. Social roles and network structure reshape how individual agents revise their beliefs (plasticity, directedness) but leave where the population converges nearly unchanged. 2. Can the resulting dynamics be explained by classical opinion-dynamics models? Only partially: the mechanisms that best predict individual belief transitions (agent identity, local neighborhood composition) are not necessarily the same as those that reproduce collective consensus, so agent- and population-level predictability diverge. Realistic simulation of belief diffusion in multi-agent LLM systems requires a diverse set of underlying LLMs. Persona prompting alone does not produce variation in collective beliefs. Limitations Scope of the experiments. Our study deliberately trades breadth for control, and several design choices limit the generalizability of our conclusions. Statements are chosen from a single domain (affirmative medical indications), which keeps the ground truth verifiable but leaves open how dynamics would change for contested, vague, or compositional claims. The LLM agent pool is narrow by design to better control for confounding factors such as family or scale. Similarly, our study does not separate the effects of finetuning data from those of optimization choices. We examine small populations (n=48n=48) over a short timeline (T=10T=10 rounds), which are large enough to observe rapid belief revisions (or changes in consensus) but insufficient to capture slow drifts or behavior at realistic community scales. No human participants were involved. Embedding humans into the same framework and comparing their update behavior with that of LLM agents under identical conditions is a natural next step. Interaction protocol. We deliberately use minimal update rules. Our protocol is synchronous, Markovian, and broadcast-style. Each LLM agent conditions only on the neighborhood summary from the previous round, not the entire interaction history. This design isolates the effects of social influence from memory and dialogue. However, it excludes practical mechanisms such as asynchronous turn-taking, persistent context, free-form messaging, and persuasion effects. Belief inference. We read beliefs from next-token probabilities over a small discrete label set. This produces a tractable, comparable signal across heterogeneous models, but it cannot characterize what happens inside an LLM (its inner representations) or address its sensitivity to prompt structure. Representation-based probes of latent truth or epistemic stability would complement this output view. These could help disentangle genuine belief revision from response conformity. Causality scope. Our causal claims are similarly bounded: the dynamics analysis detects social role, specialization, role–specialization alignment, and composition effects across the scenarios. However, we do not use formal mediation analysis or targeted interventions to localize the mechanism. Ethical Considerations CoevolveSim is a simulation of LLM-agent belief dynamics, not a study of human subjects. Therefore, the results should not be interpreted as evidence about human behavior or used as a basis for human-facing policy claims. Our findings are limited to a single Llama-3 family and may not generalize to other model families, sizes, languages, or cultural settings. Use of AI. We used Claude Code and GitHub Copilot for code annotations and formatting assistance and Grammarly for grammar and style support. After drafting the manuscript, we used Claude Fable 5 and GPT-5.6-sol to proofread it, incorporating their feedback where appropriate. Author Contributions Conceptualization: TER (lead, framing & original idea: network-based coevolution of LLM beliefs), GS (lead, experimental paradigm; supporting, framing & initial feasibility exploration), SD (lead, experimental paradigm; supporting, framing & research questions), CM (lead, experimental paradigm; supporting, framing) Data Curation: GS (lead, maximin selection, statement selection, network realization), SD (lead, surrogate data quality) Formal Analysis: GS (lead, descriptive analysis, mixed-effects models, ICC, contrast analysis; supporting, exploratory analysis, surrogate analysis), SD (lead, surrogate analysis; supporting, descriptive & exploratory analysis), CM (supporting, surrogate analysis, descriptive & exploratory analysis) Funding Acquisition: TER Investigation: GS (lead, analysis, simulations), SD (lead, analysis; supporting, simulations), CM (supporting, simulations) Methodology: GS (lead, experimental scenarios & evaluation metric, mixed-effects models, contrasts, variance decomposition; supporting, framing, surrogate models), SD (lead, framing, experimental scenarios & evaluation metric, surrogate models), TER (lead, framing), CM (supporting, framing, experimental scenarios & evaluation metric) Project Administration: TER Resources: TER Software: SD (lead, simulation logic, network generation, surrogate models; responsible for the initial codebase), GS (lead, logging, scaling, configuration & setup, descriptive/statistical analysis, GitHub maintenance; supporting, simulation logic, evaluation metrics), CM (supporting, network generation, evaluation metrics, logging) Supervision: TER (lead), GS (supporting) Validation: GS Visualization: GS Writing – Original Draft: GS (lead, introduction, methods & setup, results & descriptive analysis, limitation & ethics; supporting, discussion), SD (lead, discussion, results & surrogate models; supporting, introduction, methods & setup, literature review, discussion, results & descriptive analysis), CM (lead, literature review) Writing – Review & Editing: GS (lead), TER (lead), SD (supporting) References A. Asai, J. He, R. Shao, W. Shi, A. Singh, J. C. Chang, K. Lo, L. Soldaini, S. Feldman, M. D’Arcy, D. Wadden, M. Latzke, J. Sparks, J. D. Hwang, V. Kishore, M. Tian, P. Ji, S. Liu, H. Tong, B. Wu, Y. Xiong, L. Zettlemoyer, G. Neubig, D. S. Weld, D. Downey, W. Yih, P. W. Koh, and H. Hajishirzi (2026) Synthesizing scientific literature with retrieval-augmented language models. Nature 650 (8103), p. 857–863. Note: Hugging Face checkpoint: OpenSciLM/Llama-3.1_OpenScholar-8B External Links: ISSN 1476-4687, Link, Document Cited by: Table A2. P. E. Bailey, T. Leon, N. C. Ebner, A. A. Moustafa, and G. Weidemann (2023) A meta-analysis of the weight of advice in decision-making. Current Psychology 42 (28), p. 24516–24541. External Links: Document Cited by: §1. E. Bakshy, J. M. Hofman, W. A. Mason, and D. J. Watts (2011) Everyone’s an influencer: Quantifying influence on Twitter. In Proceedings of the Fourth ACM International Conference on Web Search and Data Mining, p. 65–74. External Links: Link, Document Cited by: §2.3, §6. A. Barabási and M. Pósfai (2016) Network science. Cambridge University Press, Cambridge. External Links: Link Cited by: §4.3. D. Bates, M. Mächler, B. Bolker, and S. Walker (2015) Fitting linear mixed-effects models using lme4. Journal of Statistical Software 67 (1). External Links: ISSN 1548-7660, Link, Document Cited by: Appendix D. A. Bawase (2024) Code Llama 3 8B. Hugging Face. Note: Hugging Face checkpoint: ajibawa-2023/Code-Llama-3-8B External Links: Link Cited by: Table A2. J. Becker, D. Brackbill, and D. Centola (2017) Network dynamics of social influence in the wisdom of crowds. Proceedings of the National Academy of Sciences 114 (26), p. E5070–E5076. External Links: Document Cited by: §1, §2.3, §6. P. Bhandari, N. Fay, M. J. Wise, A. Datta, S. Meek, U. Naseem, and M. Nasim (2025) Can LLM agents maintain a persona in discourse?. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, C. Christodoulopoulos, T. Chakraborty, C. Rose, and V. Peng (Eds.), Suzhou, China, p. 29213–29229. External Links: Link, Document, ISBN 979-8-89176-332-6 Cited by: §2.2. O. Bilgin, A. A. Sami, S. S. Vujjini, and J. Licato (2025) The effect of belief boxes and open-mindedness on persuasion. arXiv. External Links: Document, Link Cited by: §2.1. S. M. Breum, D. V. Egdal, V. Gram Mortensen, A. G. Møller, and L. M. Aiello (2024) The persuasive power of large language models. Proceedings of the International AAAI Conference on Web and Social Media 18, p. 152–163. External Links: ISSN 2162-3449, Link, Document Cited by: §2.2. V. C. Brockers, D. A. Ehrlich, and V. Priesemann (2025) Disentangling interaction and bias effects in opinion dynamics of large language models. arXiv. External Links: Document, Link Cited by: §2.1. E. Y. Çalık and T. R. Akkuş (2025) Enhancing human-like responses in large language models. arXiv. Note: Hugging Face checkpoint: HumanLLMs/Human-Like-LLama3-8B-Instruct External Links: Document, Link Cited by: Table A2. B. Chen, C. Shu, E. Shareghi, N. Collier, K. Narasimhan, and S. Yao (2023) FireAct: Toward language agent fine-tuning. arXiv. External Links: Document, Link Cited by: §2.2. W. Chen, Y. Su, J. Zuo, C. Yang, C. Yuan, C. Chan, H. Yu, Y. Lu, Y. Hung, C. Qian, Y. Qin, X. Cong, R. Xie, Z. Liu, M. Sun, and J. Zhou (2024) AgentVerse: Facilitating multi-agent collaboration and exploring emergent behaviors. In The Twelfth International Conference on Learning Representations (ICLR), External Links: Link Cited by: §2.1. D. Chicco and G. Jurman (2020) The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation. BMC Genomics 21 (1), p. 6. External Links: Document Cited by: §4.5.2. C. Christophe, P. K. Kanithi, T. Raha, S. Khan, and M. A. Pimentel (2024) Med42-v2: A suite of clinical LLMs. Note: Hugging Face checkpoint: m42-health/Llama3-Med42-8B External Links: Link, Document Cited by: Table A2, §3.1, footnote 1. Y. Chuang, A. Goyal, N. Harlalka, S. Suresh, R. Hawkins, S. Yang, D. Shah, J. Hu, and T. Rogers (2024) Simulating opinion dynamics with networks of LLM-based agents. In Findings of the Association for Computational Linguistics: NAACL 2024, K. Duh, H. Gomez, and S. Bethard (Eds.), Mexico City, Mexico, p. 3326–3346. External Links: Link, Document Cited by: §2.2, §2.3, §6. Contact Doctor (2024) ContactDoctor-Bio-Medical-MultiModal-Llama-3-8B-V1: A high-performance biomedical multimodal LLM. Note: Hugging Face checkpoint: ContactDoctor/Medical-MultiModal-Llama-3-8B-V1 External Links: Link Cited by: Table A2. W. Dai, Y. Chen, Z. Dai, Y. Liu, Z. Huang, Y. Pan, B. Song, C. Zhong, X. Li, Z. Wang, Z. Feng, and Y. Zhou (2024) KALE-LM-Chem: Vision and practice toward an AI brain for chemistry. arXiv. Note: Hugging Face checkpoint: Llama3-KALE-LM-Chem-1.5-8B External Links: Document, Link Cited by: Table A2. M. H. DeGroot (1974) Reaching a consensus. Journal of the American Statistical Association 69 (345), p. 118–121. External Links: ISSN 1537-274X, Link, Document Cited by: §2.3, §4.5.2. P. Devine (2024) Tagengo: A multilingual chat dataset. arXiv. Note: Hugging Face checkpoint: lightblue/suzume-llama-3-8B-multilingual External Links: Document, Link Cited by: Table A2. P. Erdős and A. Rényi (1959) On random graphs I. Publicationes Mathematicae 6, p. 290–297. External Links: Link Cited by: §4.3. A. Estornell and Y. Liu (2024) Multi-LLM debate: Framework, principals, and interventions. In Proceedings of the 38th International Conference on Neural Information Processing Systems (NeurIPS), Vol. 37, p. 28938–28964. External Links: Link Cited by: §2.2. FDTN AI (2024) Foundation-Sec 8B Instruct. Note: Hugging Face checkpoint: fdtn-ai/Foundation-Sec-8B-Instruct External Links: Link Cited by: Table A2. N. E. Friedkin and E. C. Johnsen (1990) Social influence and opinions. The Journal of Mathematical Sociology 15 (3-4), p. 193–206. External Links: ISSN 1545-5874, Link, Document Cited by: §2.3, §4.5.2. V. Gallego (2024) Roleplay Llama 3 8B. Hugging Face. Note: Hugging Face checkpoint: vicgalle/Roleplay-Llama-3-8B External Links: Link Cited by: Table A2. C. Gao, X. Lan, Z. Lu, J. Mao, J. Piao, H. Wang, D. Jin, and Y. Li (2023) S3: Social-network simulation system with large language model-empowered agents. arXiv. External Links: Document, Link Cited by: §2.3. T. F. Gonzalez (1985) Clustering to minimize the maximum intercluster distance. Theoretical Computer Science 38, p. 293–306. External Links: ISSN 0304-3975, Link, Document Cited by: §C.1. A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan, A. Yang, A. Fan, A. Goyal, A. Hartshorn, A. Yang, A. Mitra, A. Sravankumar, A. Korenev, A. Hinsvark, A. Rao, A. Zhang, A. Rodriguez, A. Gregerson, A. Spataru, B. Roziere, B. Biron, B. Tang, B. Chern, C. Caucheteux, C. Nayak, C. Bi, C. Marra, C. McConnell, C. Keller, C. Touret, C. Wu, C. Wong, C. C. Ferrer, C. Nikolaidis, D. Allonsius, D. Song, D. Pintz, D. Livshits, D. Wyatt, D. Esiobu, D. Choudhary, D. Mahajan, D. Garcia-Olano, D. Perino, D. Hupkes, E. Lakomkin, E. AlBadawy, E. Lobanova, E. Dinan, E. M. Smith, F. Radenovic, F. Guzmán, F. Zhang, G. Synnaeve, G. Lee, G. L. Anderson, G. Thattai, G. Nail, G. Mialon, G. Pang, G. Cucurell, H. Nguyen, H. Korevaar, H. Xu, H. Touvron, I. Zarov, I. A. Ibarra, I. Kloumann, I. Misra, I. Evtimov, J. Zhang, J. Copet, J. Lee, J. Geffert, J. Vranes, J. Park, J. Mahadeokar, J. Shah, J. van der Linde, J. Billock, J. Hong, J. Lee, J. Fu, J. Chi, J. Huang, J. Liu, J. Wang, J. Yu, J. Bitton, J. Spisak, J. Park, J. Rocca, J. Johnstun, J. Saxe, J. Jia, K. V. Alwala, K. Prasad, K. Upasani, K. Plawiak, K. Li, K. Heafield, K. Stone, K. El-Arini, K. Iyer, K. Malik, K. Chiu, K. Bhalla, K. Lakhotia, L. Rantala-Yeary, L. van der Maaten, L. Chen, L. Tan, L. Jenkins, L. Martin, L. Madaan, L. Malo, L. Blecher, L. Landzaat, L. de Oliveira, M. Muzzi, M. Pasupuleti, M. Singh, M. Paluri, M. Kardas, M. Tsimpoukelli, M. Oldham, M. Rita, M. Pavlova, M. Kambadur, M. Lewis, M. Si, M. K. Singh, M. Hassan, N. Goyal, N. Torabi, N. Bashlykov, N. Bogoychev, N. Chatterji, N. Zhang, O. Duchenne, O. Çelebi, P. Alrassy, P. Zhang, P. Li, P. Vasic, P. Weng, P. Bhargava, P. Dubal, P. Krishnan, P. S. Koura, P. Xu, Q. He, Q. Dong, R. Srinivasan, R. Ganapathy, R. Calderer, R. S. Cabral, R. Stojnic, R. Raileanu, R. Maheswari, R. Girdhar, R. Patel, R. Sauvestre, R. Polidoro, R. Sumbaly, R. Taylor, R. Silva, R. Hou, R. Wang, S. Hosseini, S. Chennabasappa, S. Singh, S. Bell, S. S. Kim, S. Edunov, S. Nie, S. Narang, S. Raparthy, S. Shen, S. Wan, S. Bhosale, S. Zhang, S. Vandenhende, S. Batra, S. Whitman, S. Sootla, S. Collot, S. Gururangan, S. Borodinsky, T. Herman, T. Fowler, T. Sheasha, T. Georgiou, T. Scialom, T. Speckbacher, T. Mihaylov, T. Xiao, U. Karn, V. Goswami, V. Gupta, V. Ramanathan, V. Kerkez, V. Gonguet, V. Do, V. Vogeti, V. Albiero, V. Petrovic, W. Chu, W. Xiong, W. Fu, W. Meers, X. Martinet, X. Wang, X. Wang, X. E. Tan, X. Xia, X. Xie, X. Jia, X. Wang, Y. Goldschlag, Y. Gaur, Y. Babaei, Y. Wen, Y. Song, Y. Zhang, Y. Li, Y. Mao, Z. D. Coudert, Z. Yan, Z. Chen, Z. Papakipos, A. Singh, A. Srivastava, A. Jain, A. Kelsey, A. Shajnfeld, A. Gangidi, A. Victoria, A. Goldstand, A. Menon, A. Sharma, A. Boesenberg, A. Baevski, A. Feinstein, A. Kallet, A. Sangani, A. Teo, A. Yunus, A. Lupu, A. Alvarado, A. Caples, A. Gu, A. Ho, A. Poulton, A. Ryan, A. Ramchandani, A. Dong, A. Franco, A. Goyal, A. Saraf, A. Chowdhury, A. Gabriel, A. Bharambe, A. Eisenman, A. Yazdan, B. James, B. Maurer, B. Leonhardi, B. Huang, B. Loyd, B. D. Paola, B. Paranjape, B. Liu, B. Wu, B. Ni, B. Hancock, B. Wasti, B. Spence, B. Stojkovic, B. Gamido, B. Montalvo, C. Parker, C. Burton, C. Mejia, C. Liu, C. Wang, C. Kim, C. Zhou, C. Hu, C. Chu, C. Cai, C. Tindal, C. Feichtenhofer, C. Gao, D. Civin, D. Beaty, D. Kreymer, D. Li, D. Adkins, D. Xu, D. Testuggine, D. David, D. Parikh, D. Liskovich, D. Foss, D. Wang, D. Le, D. Holland, E. Dowling, E. Jamil, E. Montgomery, E. Presani, E. Hahn, E. Wood, E. Le, E. Brinkman, E. Arcaute, E. Dunbar, E. Smothers, F. Sun, F. Kreuk, F. Tian, F. Kokkinos, F. Ozgenel, F. Caggioni, F. Kanayet, F. Seide, G. M. Florez, G. Schwarz, G. Badeer, G. Swee, G. Halpern, G. Herman, G. Sizov, Guangyi, Zhang, G. Lakshminarayanan, H. Inan, H. Shojanazeri, H. Zou, H. Wang, H. Zha, H. Habeeb, H. Rudolph, H. Suk, H. Aspegren, H. Goldman, H. Zhan, I. Damlaj, I. Molybog, I. Tufanov, I. Leontiadis, I. Veliche, I. Gat, J. Weissman, J. Geboski, J. Kohli, J. Lam, J. Asher, J. Gaya, J. Marcus, J. Tang, J. Chan, J. Zhen, J. Reizenstein, J. Teboul, J. Zhong, J. Jin, J. Yang, J. Cummings, J. Carvill, J. Shepard, J. McPhie, J. Torres, J. Ginsburg, J. Wang, K. Wu, K. H. U, K. Saxena, K. Khandelwal, K. Zand, K. Matosich, K. Veeraraghavan, K. Michelena, K. Li, K. Jagadeesh, K. Huang, K. Chawla, K. Huang, L. Chen, L. Garg, L. A, L. Silva, L. Bell, L. Zhang, L. Guo, L. Yu, L. Moshkovich, L. Wehrstedt, M. Khabsa, M. Avalani, M. Bhatt, M. Mankus, M. Hasson, M. Lennie, M. Reso, M. Groshev, M. Naumov, M. Lathi, M. Keneally, M. Liu, M. L. Seltzer, M. Valko, M. Restrepo, M. Patel, M. Vyatskov, M. Samvelyan, M. Clark, M. Macey, M. Wang, M. J. Hermoso, M. Metanat, M. Rastegari, M. Bansal, N. Santhanam, N. Parks, N. White, N. Bawa, N. Singhal, N. Egebo, N. Usunier, N. Mehta, N. P. Laptev, N. Dong, N. Cheng, O. Chernoguz, O. Hart, O. Salpekar, O. Kalinli, P. Kent, P. Parekh, P. Saab, P. Balaji, P. Rittner, P. Bontrager, P. Roux, P. Dollar, P. Zvyagina, P. Ratanchandani, P. Yuvraj, Q. Liang, R. Alao, R. Rodriguez, R. Ayub, R. Murthy, R. Nayani, R. Mitra, R. Parthasarathy, R. Li, R. Hogan, R. Battey, R. Wang, R. Howes, R. Rinott, S. Mehta, S. Siby, S. J. Bondu, S. Datta, S. Chugh, S. Hunt, S. Dhillon, S. Sidorov, S. Pan, S. Mahajan, S. Verma, S. Yamamoto, S. Ramaswamy, S. Lindsay, S. Lindsay, S. Feng, S. Lin, S. C. Zha, S. Patil, S. Shankar, S. Zhang, S. Zhang, S. Wang, S. Agarwal, S. Sajuyigbe, S. Chintala, S. Max, S. Chen, S. Kehoe, S. Satterfield, S. Govindaprasad, S. Gupta, S. Deng, S. Cho, S. Virk, S. Subramanian, S. Choudhury, S. Goldman, T. Remez, T. Glaser, T. Best, T. Koehler, T. Robinson, T. Li, T. Zhang, T. Matthews, T. Chou, T. Shaked, V. Vontimitta, V. Ajayi, V. Montanez, V. Mohan, V. S. Kumar, V. Mangla, V. Ionescu, V. Poenaru, V. T. Mihailescu, V. Ivanov, W. Li, W. Wang, W. Jiang, W. Bouaziz, W. Constable, X. Tang, X. Wu, X. Wang, X. Wu, X. Gao, Y. Kleinman, Y. Chen, Y. Hu, Y. Jia, Y. Qi, Y. Li, Y. Zhang, Y. Zhang, Y. Adi, Y. Nam, Yu, Wang, Y. Zhao, Y. Hao, Y. Qian, Y. Li, Y. He, Z. Rait, Z. DeVito, Z. Rosnbrick, Z. Wen, Z. Yang, Z. Zhao, and Z. Ma (2024) The Llama 3 herd of models. arXiv. Note: Hugging Face checkpoints: meta-llama/Llama-3.1-8B-Instruct and meta-llama/Meta-Llama-3-8B-Instruct External Links: Document, Link Cited by: item 1, Table A2, §4.1. T. Guo, X. Chen, Y. Wang, R. Chang, S. Pei, N. V. Chawla, O. Wiest, and X. Zhang (2024) Large language model based multi-agents: A survey of progress and challenges. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, K. Larson (Ed.), p. 8048–8057. Note: Survey Track External Links: Document, Link Cited by: §1, §1, §2.1. M. E. Johnson, L. M. Moore, and D. Ylvisaker (1990) Minimax and maximin distance designs. Journal of Statistical Planning and Inference 26 (2), p. 131–148. External Links: ISSN 0378-3758, Link, Document Cited by: Appendix C, §4.2. A. Kuznetsova, P. B. Brockhoff, and R. H. B. Christensen (2017) lmerTest package: Tests in linear mixed effects models. Journal of Statistical Software 82 (13). External Links: ISSN 1548-7660, Link, Document Cited by: §D.3. R. V. Lenth and J. Piaskowski (2026) emmeans: Estimated marginal means, aka least-squares means. Note: R package version 2.0.4 External Links: Link Cited by: §D.3. J. Leskovec (2017) Small-World Phenomenon. Note: Lecture slides from CS224W: Machine Learning with Graphs, Stanford University External Links: Link Cited by: §4.3. W. Liang, Y. Zhang, M. Codreanu, J. Wang, H. Cao, and J. Zou (2025a) The widespread adoption of large language model-assisted writing across society. Patterns 6 (12). External Links: Document Cited by: §1. W. Liang, Y. Zhang, Z. Wu, H. Lepp, W. Ji, X. Zhao, H. Cao, S. Liu, S. He, Z. Huang, D. Yang, C. Potts, C. D. Manning, and J. Zou (2025b) Quantifying large language model usage in scientific papers. Nature Human Behaviour, p. 1–11. External Links: Document Cited by: §1. J. Lorenz, H. Rauhut, F. Schweitzer, and D. Helbing (2011) How social influence can undermine the wisdom of crowd effect. Proceedings of the National Academy of Sciences 108 (22), p. 9020–9025. External Links: Document Cited by: §1, §2.3, §6. McKinsey & Company (2025) The state of AI: How organizations are rewiring to capture value. Note: McKinsey Global Institute Report External Links: Link Cited by: §1. M. Moussaïd, J. E. Kämmer, P. P. Analytis, and H. Neth (2013) Social influence and the collective dynamics of opinion formation. PLoS ONE 8 (11), p. e78433. External Links: Document Cited by: §1, §2.3, §6. mukaj (2024) Llama 3.1 Hawkish 8B. Note: Hugging Face checkpoint: mukaj/Llama-3.1-Hawkish-8B External Links: Link Cited by: Table A2. A. Mullick, S. Bose, R. Saha, A. Bhowmick, P. Goyal, N. Ganguly, P. Dey, and R. Kokku (2024) On the persona-based summarization of domain-specific documents. In Findings of the Association for Computational Linguistics: ACL 2024, L. Ku, A. Martins, and V. Srikumar (Eds.), Bangkok, Thailand, p. 14291–14307. External Links: Link, Document Cited by: §2.2. T. Naous, P. Laban, W. Xu, and J. Neville (2026) Flipping the dialogue: Training and evaluating user language models. In The Fourteenth International Conference on Learning Representations (ICLR), Note: Hugging Face checkpoint: microsoft/UserLM-8B External Links: Link Cited by: Table A2. L. H. X. Ng, I. J. Cruickshank, A. X. W. Lim, and K. M. Carley (2026) Social theory should be a structural prior for agentic AI: A formal framework for multi-agent social systems. arXiv. External Links: Document, Link Cited by: §1, §2.1. A. Pal, A. Liang, T. Kitanovski, A. Potti, and M. Goldblum (2025) Do large language models defend their beliefs consistently?. In First Workshop on Multi-Turn Interactions in Large Language Models, External Links: Link Cited by: §2.1. F. Periti, D. Alfter, and N. Tahmasebi (2024) Automatically generated definitions and their utility for modeling word meaning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Y. Al-Onaizan, M. Bansal, and Y. Chen (Eds.), Miami, Florida, USA, p. 14008–14026. External Links: Link, Document Cited by: Table A2. G. Savcisens and T. Eliassi-Rad (2025) Trilemma of truth in large language models. In Mechanistic Interpretability Workshop at NeurIPS 2025, External Links: Link Cited by: §C.3, §3.3, §4.2. F. Scheipl, S. Greven, and H. Küchenhoff (2008) Size and power of tests for a zero random effect variance or polynomial regression in additive and linear mixed models. Computational Statistics & Data Analysis 52, p. 3283–3299. External Links: Link Cited by: §D.2, §4.5.1. Z. Sun, Z. Zhang, X. Shen, Z. Zhang, Y. Liu, M. Backes, Y. Zhang, and X. He (2025) Are we in the AI-generated text world already? quantifying and monitoring AIGT on social media. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar (Eds.), Vienna, Austria, p. 22975–23005. External Links: Link, Document, ISBN 979-8-89176-251-0 Cited by: §1. E. Tanwar, D. Nathani, W. Y. Wang, and T. Chakraborty (2025) Understanding the effects of domain finetuning on LLMs. arXiv. External Links: Document, Link Cited by: §2.2. A. Taubenfeld, Y. Dover, R. Reichart, and A. Goldstein (2024) Systematic biases in LLM simulations of debates. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Y. Al-Onaizan, M. Bansal, and Y. Chen (Eds.), Miami, Florida, USA, p. 251–267. External Links: Link, Document Cited by: §2.2, §6. R. Teknium, J. Quesnelle, and C. Guang (2024) Hermes 3 technical report. arXiv. Note: Hugging Face checkpoint: NousResearch/Hermes-3-Llama-3.1-8B External Links: Document, Link Cited by: Table A2. S. Toshniwal, W. Du, I. Moshkov, B. Kisacanin, A. Ayrapetyan, and I. Gitman (2024) OpenMathInstruct-2: Accelerating AI for math with massive open-source instruction data. arXiv. Note: Hugging Face checkpoint: nvidia/OpenMath2-Llama3.1-8B External Links: Document, Link Cited by: Table A2. K. Tran, D. Dao, M. Nguyen, Q. Pham, B. O’Sullivan, and H. D. Nguyen (2025) Multi-agent collaboration mechanisms: A survey of LLMs. arXiv. External Links: Document Cited by: §1. L. Villa, D. Carneros-Prado, C. C. Dobrescu, A. Sánchez-Miguel, G. Cubero, and R. Hervás (2024) Comparative analysis of generic and fine-tuned large language models for conversational agent systems. Robotics 13 (5), p. 68. External Links: ISSN 2218-6581, Link, Document Cited by: §2.2. D. J. Watts and S. H. Strogatz (1998) Collective dynamics of ‘small-world’ networks. Nature 393 (6684), p. 440–442. External Links: ISSN 1476-4687, Link, Document Cited by: §4.3. J. Westfall, D. A. Kenny, and C. M. Judd (2014) Statistical power and optimal design in experiments in which samples of participants respond to samples of stimuli. Journal of Experimental Psychology: General 143 (5), p. 2020–2045. External Links: ISSN 0096-3445, Link, Document Cited by: §D.3. Q. Wu, G. Bansal, J. Zhang, Y. Wu, B. Li, E. Zhu, L. Jiang, X. Zhang, S. Zhang, J. Liu, A. H. Awadallah, R. W. White, D. Burger, and C. Wang (2024) AutoGen: Enabling next-gen LLM applications via multi-agent conversations. In First Conference on Language Modeling, External Links: Link Cited by: §1, §2.1. K. Xiong, X. Ding, Y. Cao, T. Liu, and B. Qin (2023) Examining inter-consistency of large language models collaboration: an in-depth analysis via debate. In Findings of the Association for Computational Linguistics: EMNLP 2023, H. Bouamor, J. Pino, and K. Bali (Eds.), Singapore, p. 7572–7590. External Links: Link, Document Cited by: §2.2. J. Ye, L. Cao, D. Chen, and E. Ferrara (2026) Stop drawing scientific claims from LLM social simulations without robustness audits. arXiv. External Links: Document, Link Cited by: §1. Appendix Appendix A Notations Symbol Description Notes ℳM Large language model (LLM) V Vocabulary of model ℳM, consisting of tokens [τ1,…,τ||]∈[ _1,…, _|V|] τ Token from the vocabulary ℳV_M e.g. tokens [1], [2] ℳ(τ∣)P_M (τ x ) Output of ℳM: a conditional probability distribution over tokens Agents and Networks G(n,p)G(n,p) Erdős–Rényi (ER) network structure G(n,β,k)G(n,β,k) Watts–Strogatz (WS) network structure n Number of nodes in a network G p Probability of an edge between two nodes in an ER network β Edge rewiring probability in a WS network k Initial degree (i.e., number of nearest neighbors) in the WS ring lattice aia_i i-th agent/node in a network G (i)N(i) Neighbors of a node i Belief i(t) b^(t)_i (Uncalibrated) belief scores of a model ℳM at round t ∑yℳ(t)(y|input)=1 _y b_M^(t)(y|input)=1 Δi(t) b_i^(t) Agent aia_i’s belief change between rounds t and t−1t-1 Δi(t)=i(t)−i(t−1) b_i^(t)= b_i^(t)- b_i^(t-1) δi(t) _i^(t) Magnitude of the per-round belief change δi(t):=12‖Δi(t)‖1 _i^(t):= 12 \| b_i^(t) \|_1 Bi(t)B_i^(t) Discretized belief label of the i-th agent Bi(t)=arg⁡maxy⁡i(y∣i(t))B_i^(t)= _y b_i(y _i^(t)) B^i(t+1) B_i^(t+1) Predicted next-round belief for agent aia_i T Total number of rounds T=10T=10 r Simulation run index c Contrast between two scenarios c=EMM1−EMM2c=EMM_1-EMM_2 Δc c Interaction contrast between two networks Δc=cER−cWS c=c_ER-c_WS d Cohen’s effect size Section˜D.3 Inputs and Text i(t)m_i^(t) Message to i-th agent at round t i(t)=(r,s,n(t))m_i^(t)=(x_r,x_s,x_n^(t)) rx_r Role specification Table˜A2 sx_s Statement (topic of discussion) Table˜A4 n(t)x_n^(t) Summary of neighbors’ beliefs at round t n(0)=∅x_n^(0)= y Ground-truth label for the statement sx_s Surrogate (Opinion-dynamics) Models PrPr Probability given by surrogate models Γ(r,t) ^(r,t) Global belief-state composition at a round t in a run r (f1(r,t),f2(r,t),f3(r,t))(f_1^(r,t),f_2^(r,t),f_3^(r,t)) Ni(r,t)N_i^(r,t) Neighborhood belief-state composition for an agent aia_i (ni,1(r,t),ni,2(r,t),ni,3(r,t))(n_i,1^(r,t),n_i,2^(r,t),n_i,3^(r,t)) MiM_i LLM assigned to an agent aia_i RiR_i Role assigned to an agent aia_i ϕi(t) φ_i^(t) Feature vector used to predict the next-round belief of agent aia_i Table A1: Notations used throughout the paper. Symbols are grouped by category. Appendix B Selection of Large Language Models We select large language models (LLMs) according to the following criteria: 1. The finetuned LLM must be based on Llama-3-Instruct or Llama-3.1-Instruct Grattafiori et al. (2024). 2. The finetuning must use conversational or instruction-based datasets to maintain the LLM’s chat-like capabilities. 3. The LLM must be publicly available on the Hugging Face platform. 4. The developers must provide documentation describing the data sources (e.g., URLs and data cards). 5. The model card or associated manuscript must report benchmark results demonstrating that the finetuning yields comparable or improved performance relative to the chat-based Llama-3 model (for a domain-specific task). Model Name Specialization (Domain) Social Role Developer & Source Llama-3.1-8B-Instruct Generalist (Base Model) LLM or Human Participant Grattafiori et al. (2024) Bio-Medical-Llama-3-8B Biomedical Research Biomedical Researcher Contact Doctor (2024) Llama3-Med42-8B Clinical Medicine Clinical Physician Christophe et al. (2024) Llama-3.1 OpenScholar-8B Scientific Literature Academic Scholar Asai et al. (2026) Llama3Dictionary-Merge Lexicography / Definitions Lexicographer Periti et al. (2024) Llama-3.1-Hawkish-8B Financial Markets Financial Analyst mukaj (2024) Llama3-KALE-LM-Chem-1.5-8B Chemistry Chemist Dai et al. (2024) Foundation-Sec-8B-Instruct Cybersecurity Security Analyst FDTN AI (2024) UserLM-8B Social Conversation Online Friend Naous et al. (2026) OpenMath2-Llama3.1-8B Mathematics Mathematician Toshniwal et al. (2024) Human-Like-Llama3-8B-Instruct Human-like Conversation Assistant Çalık and Akkuş (2025) Hermes-3-Llama-3.1-8B Strategic Reasoning Strategic Planner Teknium et al. (2024) Suzume-Llama-3-8B-Multilingual Multilingual Dialogue Language Mediator Devine (2024) Code-Llama-3-8B Coding and Maths Software Engineer Bawase (2024) Roleplay-Llama-3-8B Roleplay / Storytelling Storyteller Gallego (2024) Table A2: Summary of the 1515 LLMs used as agents in our study. Each model is characterized by its specialization (i.e., the datasets on which it was finetuned), the social role assigned to it within the experiments, and its developer or source. The selection includes both general-purpose (a.k.a. generalist) and finetuned (a.k.a. specialist) Llama-3 and Llama-3.1 models. Note: Although the generalist LLM can take the social role of a Human Participant, this designation refers only to its social identity in the experiment and does not imply human-like qualities; in scenario I, we only use the “LLM” role. In scenarios I and I, social roles are assigned at random and thus do not necessarily correspond to the specialization categories shown here. Appendix C Sampling with Maximin Criterion The design space of our experiments requires around 10510^5 simulations (statements × network realizations × scenarios × network structures), which is infeasible given our time and resource constraints. To reduce the number of simulations while ensuring that we select a diverse set of statements and network realizations, we use the maximin criterion Johnson et al. (1990). This appendix describes the algorithm, its application, and the diagnostics we use to verify coverage. C.1 Maximin Criterion Given a candidate pool =x1,…,xN⊂ℝdX=\x_1,…,x_N\ ^d of size N, we select a subset S⋆⊆S of size K whose minimum pairwise Euclidean distance is as large as possible: S⋆ S =arg⁡maxS⊆|S|=K⁡dmin(S), =\; _ subarraycS \\ |S|=K subarrayd_ (S), (A1) dmin(S) d_ (S) :=minx,y∈Sx≠y∥x−y∥2. := _ subarraycx,y\,∈ S\\ x≠ y subarray\, x-y _2. (A2) We use the farthest-point greedy algorithm of Gonzalez (1985) to perform the iterative selection in Algorithm˜1. Procedure. We start by standardizing features of X to zero mean and unit variance. We then either select the sample with the largest norm or provide an initial set S0S_0. We iteratively select the candidates from X whose pairwise minimum distance to the already-selected set S⋆S is greatest. Algorithm 1 Greedy maximin selection 1:Pool =x1,…,xN⊂ℝdX=\x_1,…,x_N\ ^d, target size K, optional starting set S0S_0. 2:S⊆S approximating S⋆S in (A1). 3:←standardize()X← standardize(X) 4:if S0S_0 is non-empty then 5: S←S0S← S_0 6:else 7: S←argmaxx∈∥x∥2S←\\, _x x _2\,\ 8:end if 9:while |S|<K|S|<K do 10: for each x∈∖Sx S do 11: d(x)←minx′∈S∥x−x′∥2d(x)← _x ∈ S\, x-x _2 12: end for 13: S←S∪arg⁡maxx∈∖S⁡d(x)S← S∪\\, _x Sd(x)\,\ 14:end while 15:return S C.2 Sampling Network Realizations with Maximin Criterion Seed ℓ¯ C k¯ k σk _k k0k_0 Erdős–Rényi (n=48n=48, p=0.3p=0.3) 2012155723 1.755 0.233 12.375 2.833 11 2292123236 1.666 0.353 15.750 3.388 20 43019216 1.734 0.322 13.542 3.937 7 1982018573 1.697 0.295 14.458 2.273 11 3781092945 1.707 0.289 13.958 3.041 27 2701326989 1.686 0.347 15.458 3.974 9 3798279163 1.725 0.284 13.625 3.289 15 601535774 1.675 0.307 15.417 2.457 21 Watts–Strogatz (n=48n=48, k=8k=8, β=0.1β=0.1) 1769674741 2.278 0.418 8.000 1.307 10 365860106 2.615 0.573 8.000 0.500 7 395036430 2.302 0.431 8.000 1.041 5 3830640276 2.352 0.468 8.000 0.645 9 1093773836 2.456 0.497 8.000 1.137 8 1356044289 2.447 0.506 8.000 0.791 6 851350617 2.621 0.538 8.000 0.816 9 4262934205 2.277 0.408 8.000 0.979 8 Table A3: Overview of the structural features of the eight networks selected per network structure via Algorithm 1. These networks are generated using our network generators (ER and WS). Features: (ℓ¯)( ) mean shortest-path length; (C)(C) global clustering coefficient; (k¯)( k) mean degree; (σk)( _k) standard deviation of the degree distribution; (k0)(k_0) degree of node 0. We consider two network structures: Erdős–Rényi (ER) and Watts–Strogatz (WS). For each structure, we run simulations across multiple network realizations to ensure the findings are not driven by a single sample (i.e., we want a structurally diverse set of realizations for each network structure). Specifically, we generate Npool=1000N_pool=1000 candidate realizations by drawing independent seeds for the corresponding generator while holding the following parameters fixed: n=48n=48 for both network structures, p=0.3p=0.3 for ER, and k=8k=8, β=0.1β=0.1 for WS. For each sampled network realization, we compute five structural features: 1. mean shortest-path length, 2. global clustering coefficient, 3. mean degree, 4. standard deviation of the degree distribution, 5. degree of node 0, the node reserved for the Clinical Physician agent (so that the specialist’s local connectivity varies across replicates). We discard candidate network realizations with more than one connected component. The resulting per-structure feature matrices are passed to Algorithm˜1 with K=8K=8. That is, as a result of the maximin sampling, we obtain eight network realizations per network structure. Table˜A3 provides an overview of the selected network realizations. C.3 Sampling Discussion Statements with Maximin Criterion ID Statement y^doc y_doc adoca_doc aothera_other c False statements (y=0y=0) false_0 Silver is indicated for the treatment of keratosis pilaris. 0 0.999 0.800 0.948 false_1 Terbutaline is indicated for the treatment of cramps. 1 0.078 0.067 0.966 false_2 Isosorbide is indicated for the treatment of ischemia. 1 0.012 0.867 0.024 false_3 Atorvastatin is indicated for the treatment of aggression. 0 1.000 0.000 0.022 false_4 Manidipine is indicated for the treatment of septic. 0 1.000 0.333 0.554 false_5 VLTS-589 is indicated for the treatment of diabetic macular edema. 1 0.383 0.867 0.386 false_6 Nandrolone decanoate is indicated for the treatment of Duchenne muscular dystrophy. 0 1.000 0.667 0.608 false_7 Vitamin E is indicated for the treatment of pulmonary coccidioidomycosis. 0 1.000 0.267 0.191 false_8 Tetracycline is indicated for the treatment of throat. 0 0.755 0.933 0.763 false_9 Ethinylestradiol is indicated for the treatment of dry cough. 0 1.000 0.067 0.301 True statements (y=1y=1) true_0 Methyl nicotinate is indicated for the treatment of aches. 1 0.773 0.933 0.869 true_1 Terbutaline is indicated for the treatment of asthma. 0 0.000 0.133 0.986 true_2 Warfarin is indicated for the treatment of pulmonary embolism. 1 1.000 0.200 0.006 true_3 Tetracycline is indicated for the treatment of Q fever. 0 0.047 0.933 0.188 true_4 Theophylline is indicated for the treatment of asthma. 1 0.531 0.267 0.773 true_5 Sildenafil is indicated for the treatment of pulmonary hypertension. 1 1.000 0.533 0.510 true_6 Polymyxin B is indicated for the treatment of infections of the urinary tract. 0 0.438 0.133 0.571 true_7 SNS-314 is indicated for the treatment of tumors. 0 0.016 0.533 0.589 true_8 Valomaciclovir is indicated for the treatment of viral infection. 0 0.318 0.400 0.920 true_9 Linezolid is indicated for the treatment of nosocomial pneumonia. 1 0.703 0.200 0.307 Table A4: List of 20 discussion statements selected via Algorithm˜1 and their corresponding ground-truth label and four features. We divide the statements into false (y=0y=0) and true (y=1y=1) subtables. The four maximin features are the medical specialist’s predicted label (y^doc)( y_doc), the medical specialist’s accuracy (adoc)(a_doc), the mean accuracy of the other 14 LLMs (aother)(a_other), and the consensus score (c)(c) between the medical specialist and the 1414 other LLMs: 1−|pdoc(true∣s)−p~other(true∣s)|1- p_doc( true s)- p_other( true s) , where p~other p_other is the median probability of the 1414 other LLMs. Note: We do not correct the template-induced grammatical artifacts so that prompts remain uniformly formatted across conditions. Our statements are drawn from the medical-indication data introduced in Savcisens and Eliassi-Rad (2025). We restrict selection to non-negated statements that reference real-world entities.999We exclude neither-valued statements with synthetic entities. An example of such a statement is “Heparin is indicated for the treatment of pulmonary embolism.” Before applying the maximin criterion, we compute zero-shot predictions for each statement s from the medical specialist (Llama3-Med42-8B) and from 14 other LLMs in Table˜A2. In addition, we define the following features for each statement s: 1. The binary ground-truth label: y∈false,truey∈\ false, true\ or 0,1\0,1\, respectively, 2. Label predicted by a medical specialist, 3. Medical specialist’s accuracy, based on next-token probabilities: 1−|y−pdoc(true∣s)|1- y-p_doc( true s) , 4. Label-level (averaged) accuracy of other LLMs, 5. Consensus between the medical specialist and other models =1−|pdoc(true∣s)−p~other(true∣s)|=1- p_doc( true s)- p_other( true s) , where p~other p_other denotes the median probability assigned by the 14 other models. The first two features are binary; the remaining three are continuous in the range [0,1][0,1]. Note that feature #4\#4 is computed from the binary label predictions, while feature #5\#5 compares probability distributions, so the two are not redundant. Since we want to have an equal number of true and false statements, we add a constraint to Algorithm˜1 that enforces a K/2K/2 quota per ground-truth label (we set K=20K=20). That is, we ensure that we select 1010 true and 1010 false statements. Table˜A4 displays selected statements with the corresponding features. Appendix D Additional Details on Descriptive Analysis This section expands on the methodology described in §4.5. Here, we use R’s mixed-model (lme4) formula notation Bates et al. (2015). A term y ˜ x + (1|g) fits a fixed effect for x and a random intercept for each level of grouping factor g (a separate baseline offset per level of g, which accounts for repeated measures within each level). Meanwhile, (1|g1:g2) denotes a random intercept for each combination of levels of g1 and g2; it is used when g2’s labels are only meaningful jointly with g1. For instance, “network realization 3” refers to a different network under ER vs. WS. Finally, n-str stands for the network structure (ER or WS) and n-realization stands for the network realization; each network structure has eight network realizations (Table˜A3). We use two identifiers for a network realization, depending on the analysis: n-realization indexes a realization within its network structure (values 1–8, reused across ER and WS) and n-realization-id is a globally unique identifier assigned to each of the 16 realizations across both structures. D.1 Linear Mixed-Effects Models Estimated marginal means (EMMs), or least-squares means, are model-based predicted means adjusted for the other terms in the mixed model. To estimate the coefficients, EMMs, and confidence intervals, we fit the following linear mixed-effects model for plasticity, directedness, consensus (and consensus change), and outgoing influence (see §4.5.1 for definitions): <var-name> ˜ scenario * n-str + (1 | statement) + (1 | n-str:n-realization). The random-effects structure includes (1|statement) to account for repeated measures across statements and (1|n-str:n-realization) to capture variation due to different realizations of each network structure. We provide an overview of results in Table˜A6. Here, we include the marginal estimates where the mixed-effects model is specified as: <var-name> ˜ scenario + (1 | statement) + (1 | n-str:n-realization). D.2 Variance Decomposition To assess whether some agents exert more influence than others, we fit a variance‑components model to the agent‑level outgoing Influence(i)Influence(i), with intercept‑only fixed effects and random intercepts for agent (nested within networks), network realization, and statement (we fit models separately for each scenario). The model formula is influence ˜ 1 + (1 | n-realization-id) + (1 | statement) + (1 | agent:n-realization-id). It allows us to quantify the variance in outgoing influence attributed to agent identity relative to network realization and statement. For each random‑effects component, we report the intraclass correlation coefficient (ICC) as the proportion of total variance attributable to that component. A higher agent‑level ICC suggests asymmetry in the outgoing influence (presence of opinion leaders and followers), whereas a near‑zero value indicates that influence is largely exchangeable across agents (no opinion leaders). In Figure˜4, we report 95% confidence intervals estimated via a parametric bootstrap (1,0001,000 resamples). To test whether the agent‑level variance component is distinguishable from zero, we use an exact restricted likelihood ratio test (RLRsim; Scheipl et al., 2008), which accounts for the boundary problem in testing σagent2=0σ^2_agent=0. The null distribution is simulated with 10,00010,000 Monte Carlo draws. D.3 Contrast Analysis We define four planned contrasts on the four scenarios (I, I, I, and IV), each corresponding to a single‑degree‑of‑freedom effect on the outcome. These four contrasts were specified a priori, before inspecting the fitted models. We define the contrast as c=EMM1−EMM2c=EMM_1-EMM_2, where EMMs are based on the outcomes of two scenarios. Table˜A5 specifies the contrast weights, and below we provide the contrast definitions: Role Effect: Random Roles (I) vs. Baseline (I) Specialization Effect: Random Specialists (I) vs. Random Roles (I) Role–specialization (alignment) effect: Matched Specialists (IV) vs. Random Specialists (I) Composition Effect: comparing the average of homogeneous scenarios (I and I) and the average of heterogeneous scenarios (I and IV). Contrast Notation I. Baseline I. Random Roles I. Random Specialists IV. Matched Specialists Role effect I→ −1-1 +1+1 0 0 Specialization effect I→ 0 −1-1 +1+1 0 Role–specialization alignment I→ 0 0 −1-1 +1+1 Composition effect I, I→I,IV −0.5-0.5 −0.5-0.5 +0.5+0.5 +0.5+0.5 Table A5: Contrast weights for the four contrasts between scenarios. Each row defines a linear combination of the four scenarios as a single degree-of-freedom contrast. Resulting effect sizes are visualized in Figure˜A1 and reported in Tables˜A8 and A9. For each contrast above, we compute three quantities: • Marginal contrast (Table˜A8): the contrast averaged across both network structures, showing the overall effect of a manipulation. • Conditional contrast (Table˜A9): the same linear combination estimated separately within each network structure (ER or WS). • Interaction contrast (the Between rows in Table˜A9): the difference between the ER and WS conditional contrasts, cER−cWSc_ER-c_WS. A non-significant interaction indicates that we do not detect differences between manipulation effects when comparing two network structures. For all contrasts, we report Cohen’s d, 95%95\% confidence intervals, standard errors, and t-statistics. Figure A1: Effect sizes for the four planned contrasts. We provide marginal and conditional estimates for ER and WS network structures. The results are provided in Tables˜A9 and A8. Computations and effect sizes. We compute the marginal and conditional contrasts c and their 95%95\% confidence intervals from the fitted mixed-effects model with emmeans (Lenth and Piaskowski, 2026), using Satterthwaite-approximated degrees of freedom (lmerTest; Kuznetsova et al., 2017). For each contrast, we report Cohen’s d: the estimated marginal-mean (EMM) contrast c standardized by the total standard deviation of the mixed-effects model, d=cσ^total,withσ^total=σ^ϵ2+∑gτ^g2,d= c σ_total, σ_total= \, σ^2_ε+ _g τ^2_g\,, (A3) where we sum over the random-intercept variance components τ^g2 τ^2_g (e.g., statement and network realization for the agent-level outcomes) and σ^ϵ2 σ^2_ε is the residual variance (Westfall et al., 2014). In R, it is computed as the square root of the summed VarCorr components (valid for this purpose, since all random effects are intercept-only). Note that because the components entering σ^total σ_total differ by observational level, d is comparable across scenarios within an outcome but not directly across the agent and population levels. The interaction contrast (the Between rows in Table˜A9) is the difference between the ER and WS conditional contrasts, Δc=cER−cWS c=c_ER-c_WS, standardized by the same σ^total σ_total (values near zero indicate a network-invariant effect). The standard error for Δc c (SE) is obtained from the variance–covariance matrix of the EMMs. Appendix E Surrogate Model Details The surrogate analysis evaluates whether the belief dynamics observed in CoevolveSim can be reproduced using a hierarchy of transition models. E.1 Feature Construction Let Bi(r,t)∈1,2,3B_i^(r,t)∈\1,2,3\ denote the discretized belief of agent aia_i in run r at round t, where 11, 22, and 33 correspond to accepting, rejecting, or abstaining from the statement, respectively. All models condition on the agent’s current belief state Bi(r,t)B_i^(r,t). Additional features are defined as follows. Global belief composition. For run r and round t, we define Γ(r,t)=(f1(r,t),f2(r,t),f3(r,t)), ^(r,t)= (f_1^(r,t),f_2^(r,t),f_3^(r,t) ), (A4) where fc(r,t)f_c^(r,t) denotes the fraction of agents holding belief state c. Neighborhood belief composition. For agent aia_i, we define Ni(r,t)=(ni,1(r,t),ni,2(r,t),ni,3(r,t)),N_i^(r,t)= (n_i,1^(r,t),n_i,2^(r,t),n_i,3^(r,t) ), (A5) where ni,c(r,t)n_i,c^(r,t) denotes the fraction of neighbors in (i)N(i) holding belief state c. Identity features. For model M4, we additionally include MiM_i and RiR_i, the underlying LLM and social role assigned to agent aia_i, respectively. E.2 Model Specification Each surrogate estimates a conditional distribution over next-round beliefs, Pr⁡(Bi(t+1)=c∣ϕi(t)),c∈1,2,3, \! (B_i^(t+1)=c φ_i^(t) ),\;c∈\1,2,3\, (A6) where ϕi(t) φ_i^(t) is a model-specific feature vector. Models are fit separately for each experimental scenario and network structure. M1: Empirical persistence baseline. M1 estimates one-step transition probabilities directly from observed transition frequencies: Pr⁡(Bi(t+1)=c∣Bi(t)). \! (B_i^(t+1)=c B_i^(t) ). (A7) M2: Global belief composition. M2 conditions on the agent’s current belief and the global belief composition: Pr⁡(Bi(t+1)=c∣Bi(t),Γ(t)). \! (B_i^(t+1)=c B_i^(t), ^(t) ). (A8) M3: Local belief composition. M3 conditions on the agent’s current belief and local neighborhood (belief) composition: Pr⁡(Bi(t+1)=c∣Bi(t),Ni(t)). \! (B_i^(t+1)=c B_i^(t),N_i^(t) ). (A9) M4: Local belief composition + identity. M4 augments M3 with agent-level identity features: Pr⁡(Bi(t+1)=c∣Bi(t),Ni(t),Mi,Ri). \! (B_i^(t+1)=c B_i^(t),N_i^(t),M_i,R_i ). (A10) E.3 Fitting Procedure For each network structure and experimental scenario, we construct transition-level examples from the empirical belief trajectories generated by CoevolveSim. Transitions are partitioned into training, validation, and test sets at the run level so that observations from the same simulation run do not appear in multiple splits. M1 estimates transition probabilities directly from observed transition counts. Models M2–M4 are implemented as multinomial logistic regressions predicting the next-round belief state from the corresponding feature set. E.4 Evaluation We evaluate the surrogate hierarchy using both held-out prediction and behavioral rollouts. Held-out transition prediction. For each transition, the fitted model predicts B^i(t+1) B_i^(t+1) from information available at round t. Predictive performance is evaluated using the Matthews Correlation Coefficient (MCC); specifically, we look at whether the final belief state from the surrogate rollout agrees with the observed final-round state. We refer to this agreement as a final-state MCC. Rollout evaluation. For each held-out run, we initialize the surrogate from the empirical round-0 belief state and iteratively generate subsequent belief states by sampling from the fitted transition model. This procedure produces a complete surrogate belief trajectory. We compare empirical and surrogate trajectories using consensus fidelity, which measures agreement between the final-round consensus levels of the empirical and surrogate simulations. Higher consensus fidelity indicates that the surrogate more accurately reproduces the collective outcome. We provide an overview of the results in Tables˜A10 and A11. Appendix F Tables In this section, we provide detailed tables for 1. Estimated marginal means (EMMs) in Table˜A6, 2. Variance decomposition for the outgoing influence in Table˜A7, 3. Contrast analysis: marginal effect sizes in Table˜A8 and conditional effects by network structure in Table˜A9, 4. Surrogate analysis: final-state Matthews Correlation Coefficient (final-state MCC) in Table˜A10 and consensus fidelity in Table˜A11. Estimate 95%95\% CI lower 95%95\% CI upper SE df Outcome Scenario Network Plasticity I. Base LLMs ER 0.03 0.00 0.05 0.01 20.00 WS 0.03 0.00 0.06 0.01 20.00 Marginal 0.03 0.00 0.05 0.01 19.50 I. Random roles ER 0.07 0.05 0.10 0.01 20.00 WS 0.05 0.03 0.08 0.01 20.00 Marginal 0.06 0.04 0.09 0.01 19.50 I. Random Specialists ER 0.13 0.11 0.16 0.01 20.00 WS 0.11 0.09 0.14 0.01 20.00 Marginal 0.12 0.10 0.15 0.01 19.50 IV. Matched Specialists ER 0.12 0.09 0.14 0.01 20.00 WS 0.14 0.11 0.17 0.01 20.00 Marginal 0.13 0.10 0.15 0.01 19.50 Directedness I. Base LLMs ER 0.97 0.91 1.03 0.03 19.90 WS 0.96 0.90 1.02 0.03 19.90 Marginal 0.96 0.91 1.02 0.03 19.80 I. Random roles ER 0.44 0.38 0.50 0.03 19.90 WS 0.50 0.44 0.56 0.03 19.90 Marginal 0.47 0.41 0.53 0.03 19.80 I. Random Specialists ER 0.27 0.21 0.33 0.03 19.90 WS 0.35 0.29 0.41 0.03 19.90 Marginal 0.31 0.25 0.37 0.03 19.80 IV. Matched Specialists ER 0.31 0.25 0.36 0.03 19.90 WS 0.26 0.20 0.32 0.03 19.90 Marginal 0.28 0.23 0.34 0.03 19.80 Outgoing Influence I. Base LLMs ER 0.02 -0.01 0.04 0.01 19.90 WS 0.02 -0.01 0.05 0.01 19.90 Marginal 0.02 -0.01 0.05 0.01 19.40 I. Random roles ER 0.07 0.05 0.10 0.01 19.90 WS 0.05 0.03 0.08 0.01 19.90 Marginal 0.06 0.04 0.09 0.01 19.40 I. Random Specialists ER 0.13 0.11 0.16 0.01 19.90 WS 0.12 0.09 0.15 0.01 19.90 Marginal 0.13 0.10 0.16 0.01 19.40 IV. Matched Specialists ER 0.12 0.09 0.14 0.01 19.90 WS 0.15 0.12 0.17 0.01 19.90 Marginal 0.13 0.10 0.16 0.01 19.40 Consensus Change I. Base LLMs ER 0.00 -0.03 0.02 0.01 32.00 WS -0.01 -0.04 0.01 0.01 32.00 Marginal -0.01 -0.03 0.02 0.01 24.20 I. Random roles ER 0.04 0.01 0.07 0.01 32.00 WS 0.05 0.02 0.08 0.01 32.00 Marginal 0.04 0.02 0.07 0.01 24.20 I. Random Specialists ER 0.19 0.17 0.22 0.01 32.00 WS 0.18 0.15 0.21 0.01 32.00 Marginal 0.19 0.16 0.21 0.01 24.20 IV. Matched Specialists ER 0.16 0.13 0.18 0.01 32.00 WS 0.14 0.11 0.17 0.01 32.00 Marginal 0.15 0.12 0.17 0.01 24.20 Table A6: Estimated marginal means (EMMs) of agent- and population-level outcomes by scenario and network structure (Appendix˜D). EMMs with 95%95\% confidence intervals are estimated with the linear mixed-effects models, reported per network structure (ER, WS) and marginally (averaged over both structures with equal weights). Agent-level outcomes (plasticity, directedness, outgoing influence) are fit with random intercepts for statement and network realization, while consensus change is fit at the run level; df denotes Satterthwaite-approximated degrees of freedom (which is why marginal and per-network rows differ). SE is the standard error of the EMM; in the network column ER stands for Erdős–Rényi and WS for Watts–Strogatz. Agent-level ICC (95% CI) Network-level ICC (95% CI) Scenario ICC Lower Upper ICC Lower Upper RLRT p I. Baseline 0.00 0.00 0.05 0.27 0.09 0.66 0.50 I. Random Roles 0.00 0.00 0.12 1.32 0.49 2.87 0.49 I. Random Specialists 4.97 3.50 7.08 2.96 1.12 5.96 <0.001 IV. Matched Specialists 2.73 1.90 4.04 4.98 1.90 9.35 <0.001 Table A7: Variance decomposition of outgoing influence by scenario: intraclass correlations (ICC, %) show the ratio of variance in outgoing influence that is attributed to agent identities (agent-level) or network realizations (network-level). Additionally, we provide the likelihood-ratio test results (RLRT). The variation in influence due to the LLM agents’ identities (emergence of persistent opinion leaders and followers) appears only when specialist LLMs are introduced; that is, in scenarios I and IV. Cohen’s d 95%95\% CI lower 95%95\% CI upper SE t-stat p-val df Sig Outcome Contrast Plasticity Role effect 0.32 0.25 0.38 0.004 9.55 <0.001 1253 * Specialization effect 0.53 0.46 0.59 0.004 15.83 <0.001 1253 * Role–spec. alignment 0.06 -0.01 0.12 0.004 1.66 0.098 1253 Composition effect 0.71 0.67 0.76 0.003 30.30 <0.001 1253 * Directedness Role effect -1.86 -1.94 -1.79 0.010 -48.59 <0.001 1253 * Specialization effect -0.62 -0.69 -0.54 0.010 -16.14 <0.001 1253 * Role–spec. alignment -0.09 -0.17 -0.01 0.010 -2.35 0.019 1253 * Composition effect -1.60 -1.65 -1.54 0.007 -58.85 <0.001 1253 * Outgoing Influence Role effect 0.53 0.44 0.62 0.004 11.19 <0.001 1253 * Specialization effect 0.73 0.64 0.82 0.004 15.36 <0.001 1253 * Role–spec. alignment 0.04 -0.05 0.13 0.004 0.85 0.394 1253 Composition effect 1.01 0.95 1.08 0.003 30.24 <0.001 1253 * Consensus Change Role effect 0.51 0.38 0.65 0.007 7.52 <0.001 1253 * Specialization effect 1.39 1.26 1.53 0.007 20.46 <0.001 1253 * Role–spec. alignment -0.36 -0.49 -0.23 0.007 -5.27 <0.001 1253 * Composition effect 1.47 1.37 1.56 0.005 30.53 <0.001 1253 * Table A8: Marginal effects for the four contrasts on agent- and population-level outcomes (Appendix˜D). Each contrast is a single-degree-of-freedom comparison between scenarios, averaged across network structures: role effect (I vs. I), specialization effect (I vs. I), role–specialization alignment (IV vs. I), and composition effect (I, IV vs. I, I). Effect sizes are reported as Cohen’s d with 95%95\% confidence intervals; larger |d||d| indicates a larger standardized difference; the sign of d indicates the direction of the effect. SE is the standard error of the unstandardized contrast c, t-stat is the associated t-statistic, the p-value tests H0H_0: d=0d=0, df denotes Satterthwaite-approximated degrees of freedom, and ∗* in the Sig column denotes p<0.05p<0.05. Cohen’s d 95%95\% CI lower 95%95\% CI upper SE t-stat p-val df Sig Outcome Network Contrast Plasticity ER Role effect 0.43 0.34 0.52 0.005 9.09 <0.001 1253 * Specialization effect 0.53 0.44 0.62 0.005 11.24 <0.001 1253 * Role–spec. alignment -0.15 -0.24 -0.06 0.005 -3.17 0.002 1253 * Composition effect 0.67 0.60 0.74 0.004 20.09 <0.001 1253 * WS Role effect 0.21 0.12 0.30 0.005 4.41 <0.001 1253 * Specialization effect 0.53 0.43 0.62 0.005 11.14 <0.001 1253 * Role–spec. alignment 0.26 0.17 0.35 0.005 5.52 <0.001 1253 * Composition effect 0.76 0.69 0.82 0.004 22.77 <0.001 1253 * Between Role effect 0.22 0.09 0.35 0.007 3.31 <0.001 1253 * Specialization effect 0.01 -0.13 0.14 0.007 0.07 0.941 1253 Role–spec. alignment -0.41 -0.54 -0.28 0.007 -6.15 <0.001 1253 * Composition effect -0.09 -0.18 0.00 0.005 -1.90 0.058 1253 Directedness ER Role effect -1.99 -2.10 -1.89 0.014 -36.73 <0.001 1253 * Specialization effect -0.65 -0.76 -0.55 0.014 -12.03 <0.001 1253 * Role–spec. alignment 0.13 0.03 0.24 0.014 2.47 0.014 1253 * Composition effect -1.58 -1.66 -1.51 0.01 -41.24 <0.001 1253 * WS Role effect -1.74 -1.84 -1.63 0.014 -31.99 <0.001 1253 * Specialization effect -0.59 -0.69 -0.48 0.014 -10.79 <0.001 1253 * Role–spec. alignment -0.32 -0.42 -0.21 0.014 -5.80 <0.001 1253 * Composition effect -1.61 -1.69 -1.54 0.01 -41.98 <0.001 1253 * Between Role effect -0.26 -0.41 -0.11 0.020 -3.36 <0.001 1253 * Specialization effect -0.07 -0.22 0.08 0.020 -0.88 0.382 1253 Role–spec. alignment 0.45 0.30 0.60 0.020 5.84 <0.001 1253 * Composition effect 0.03 -0.08 0.14 0.014 0.52 0.602 1253 Outgoing Influence ER Role effect 0.68 0.54 0.81 0.006 10.06 <0.001 1253 * Specialization effect 0.68 0.54 0.81 0.006 10.07 <0.001 1253 * Role–spec. alignment -0.20 -0.33 -0.07 0.006 -2.96 0.003 1253 * Composition effect 0.91 0.82 1.01 0.004 19.27 <0.001 1253 * WS Role effect 0.39 0.26 0.52 0.006 5.76 <0.001 1253 * Specialization effect 0.78 0.65 0.91 0.006 11.65 <0.001 1253 * Role–spec. alignment 0.28 0.15 0.41 0.006 4.17 <0.001 1253 * Composition effect 1.11 1.02 1.21 0.004 23.50 <0.001 1253 * Between Role effect 0.29 0.10 0.47 0.008 3.04 0.002 1253 * Specialization effect -0.11 -0.29 0.08 0.008 -1.12 0.263 1253 Role–spec. alignment -0.48 -0.66 -0.29 0.008 -5.04 <0.001 1253 * Composition effect -0.20 -0.33 -0.07 0.006 -3.00 0.003 1253 * Consensus Change ER Role effect 0.43 0.24 0.62 0.010 4.44 <0.001 1253 * Specialization effect 1.50 1.31 1.69 0.010 15.63 <0.001 1253 * Role–spec. alignment -0.35 -0.54 -0.17 0.010 -3.69 <0.001 1253 * Composition effect 1.54 1.41 1.67 0.007 22.64 <0.001 1253 * WS Role effect 0.60 0.41 0.79 0.010 6.20 <0.001 1253 * Specialization effect 1.28 1.09 1.47 0.010 13.31 <0.001 1253 * Role–spec. alignment -0.36 -0.55 -0.17 0.010 -3.77 <0.001 1253 * Composition effect 1.40 1.26 1.53 0.007 20.54 <0.001 1253 * Between Role effect -0.17 -0.44 0.10 0.014 -1.25 0.213 1253 Specialization effect 0.22 -0.04 0.49 0.014 1.64 0.100 1253 Role–spec. alignment 0.01 -0.26 0.28 0.014 0.06 0.955 1253 Composition effect 0.14 -0.05 0.33 0.010 1.48 0.138 1253 Table A9: Conditional and interaction effects for the four contrasts on agent- and population-level outcomes (Appendix˜D). Each contrast is estimated separately within each network structure (ER stands for Erdős–Rényi, and WS stands for Watts–Strogatz). The Between rows report interaction contrasts (ER conditional contrast −- WS conditional contrast); a non-significant Between row indicates that the manipulation does not produce any detectable differences between the two network structures. The contrasts include role effect (I vs. I), specialization effect (I vs. I), role–specialization alignment (IV vs. I), and composition effect (I, IV vs. I, I). Effect sizes are reported as Cohen’s d with 95%95\% confidence intervals; larger |d||d| indicates a larger standardized difference; the sign of d indicates the direction of the effect. SE is the standard error of the unstandardized contrast c, t-stat is the associated t-statistic, the p-value tests H0H_0: d=0d=0, df denotes Satterthwaite-approximated degrees of freedom, and ∗* in the Sig column denotes p<0.05p<0.05. Network Final-state MCC (%) 95%95\% CI lower 95%95\% CI upper Scenario Surrogate I. Baseline M1 ER 92.4 90.0 94.5 WS 90.2 87.7 92.6 M2 ER 98.2 97.0 99.3 WS 90.2 87.7 92.6 M3 ER 97.2 95.7 98.4 WS 92.5 90.3 94.6 M4 ER 97.2 95.7 98.4 WS 92.5 90.3 94.6 I. Random Roles M1 ER 86.1 83.0 88.9 WS 94.7 92.8 96.5 M2 ER 87.5 84.7 90.2 WS 94.8 92.7 96.5 M3 ER 90.8 88.4 93.2 WS 93.8 91.8 95.8 M4 ER 90.3 87.7 92.8 WS 94.7 92.8 96.5 I. Random Specialists M1 ER 76.3 72.8 79.7 WS 84.4 81.3 87.2 M2 ER 80.7 77.6 83.9 WS 84.8 81.9 87.7 M3 ER 80.6 77.4 83.8 WS 86.4 83.6 89.1 M4 ER 82.5 79.3 85.5 WS 89.1 86.5 91.6 IV. Matched Specialists M1 ER 81.7 78.5 84.7 WS 79.6 75.9 83.1 M2 ER 79.9 76.3 83.1 WS 78.7 75.0 82.3 M3 ER 81.5 78.5 84.6 WS 80.6 77.0 83.8 M4 ER 84.7 81.7 87.6 WS 82.0 78.7 85.3 Table A10: Final-state Matthews Correlation Coefficient (final-state MCC) by scenario, surrogate model, and network structure (Appendix˜E). We report MCC with 95%95\% bootstrapped confidence intervals for agreement between each surrogate’s predicted final belief state and the observed final belief state. Surrogate models are M1 (persistence), M2 (persistence + global belief composition), M3 (persistence + local neighborhood composition), and M4 (M3 + agent identity and social role); see Table˜2. Higher final-state MCC indicates more accurate prediction of individual final-round beliefs. In the network column, ER stands for Erdős–Rényi, and WS stands for Watts–Strogatz. Network Consensus Fidelity (%) 95%95\% CI lower 95%95\% CI upper Scenario Surrogate I. Baseline M1 ER 91.1 88.7 93.6 WS 90.4 88.3 92.7 M2 ER 80.1 78.6 81.4 WS 81.7 79.6 84.1 M3 ER 81.1 78.1 84.2 WS 78.4 75.2 81.7 M4 ER 81.0 77.9 84.3 WS 78.5 75.2 81.7 I. Random Roles M1 ER 80.1 77.8 82.5 WS 82.1 79.8 84.4 M2 ER 84.7 81.7 87.7 WS 86.7 83.9 89.3 M3 ER 77.8 75.4 80.1 WS 76.0 73.8 78.1 M4 ER 77.6 75.0 80.2 WS 76.6 74.2 78.8 I. Random Specialists M1 ER 72.1 69.5 75.2 WS 73.3 70.4 76.9 M2 ER 94.5 93.1 95.6 WS 93.3 90.5 95.6 M3 ER 92.9 91.6 94.2 WS 90.9 89.0 92.6 M4 ER 93.8 92.5 94.8 WS 92.1 90.5 93.4 IV. Matched Specialists M1 ER 73.7 70.6 77.2 WS 73.4 69.6 78.2 M2 ER 93.6 91.3 95.5 WS 92.4 89.0 94.9 M3 ER 91.1 88.5 93.3 WS 90.5 88.9 91.9 M4 ER 92.5 90.6 94.1 WS 89.8 87.9 91.8 Table A11: Consensus fidelity by scenario, surrogate model, and network structure (Appendix˜E) with 95%95\% bootstrapped confidence intervals. Consensus fidelity is defined as the similarity between observed and surrogate final-round consensus levels across runs; higher values indicate that a surrogate more accurately reproduces the population-level agreement observed in CoevolveSim. Surrogate models are M1 (persistence), M2 (persistence + global belief composition), M3 (persistence + local neighborhood composition), and M4 (M3 + agent identity and social role); see Table˜2. In the network column, ER stands for Erdős–Rényi, and WS stands for Watts–Strogatz.

Instant research discovery