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

Abstract: 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.

• The paper demonstrates that LLM heterogeneity significantly drives belief revision and consensus formation more than persona-style prompts.
• It employs CoevolveSim to simulate belief updates in structured networks, contrasting generalist and domain-specialist agents.
• Findings indicate that classical opinion-dynamics models require population-level features to accurately capture heterogeneous agent behavior.

Belief Coevolution in Populations of Generalist and Specialist LLMs

Introduction

"Belief Coevolution in a Social Network of Generalist and Specialist LLMs" (2607.27512) investigates the diffusion and revision of beliefs among populations of LLM agents with varying domain specialization and social role assignments, embedded within structured social networks. The study introduces CoevolveSim, a simulation framework isolating the effects of LLM heterogeneity, persona-style prompting, and network topology on the dynamics of collective belief formation. The work systematically explores two primary questions: what factors govern the process of belief revision in LLM-based multi-agent systems, and to what extent can classical opinion-dynamics models explain these observed phenomena?

Experimental Paradigm

CoevolveSim operationalizes LLM “agents” as nodes in undirected social networks, each powered by either a generalist or a domain-specialized (finetuned) LLM. Every agent is assigned both a specialization (reflecting the LLM powering it) and a social role (a persona label communicated to itself and its neighbors). The simulation proceeds in synchronous rounds: agents receive summaries of their neighbors' belief states and revise their own beliefs about factual statements, extracted from medical indication data, according to a templated interaction protocol. Two canonical network topologies are evaluated (Erdős–Rényi and Watts–Strogatz), with careful maximin sampling of 48-node graphs to ensure structural and experimental diversity.

Four distinct simulation scenarios are constructed:

1. Homogeneous generalist LLMs with uniform roles,
2. Homogeneous generalist LLMs with randomized social roles (persona-style prompting),
3. Heterogeneous mix of generalists and specialists with random social roles,
4. Heterogeneous mix with social roles aligned (“matched”) to LLM specialization.

Each scenario is replicated across networks and discussion statements, yielding 1,280 controlled simulation runs.

Key Findings: Drivers of Belief Dynamics

LLM Heterogeneity and Social Roles

The analysis demonstrates that the introduction of LLM heterogeneity—i.e., deploying agents powered by domain-specialized LLMs—is the primary mechanism for inducing both significant belief revision at the agent level and substantial consensus shifts at the population level. Specialist LLMs more than double the shift in collective consensus compared to scenarios where diversity is induced only by persona-style prompting. Furthermore, agent-level variance decomposition reveals opinion-leader-like dynamics (persistent between-agent differences in exerted influence) only in specialist settings.

Persona-style social role assignment strongly modulates belief trajectories of individual agents, increasing update plasticity and inducing oscillations, but it has only modest impact on aggregate population-level consensus. Network topology further shapes agent-level behavior, but its effect on the final consensus state is negligible compared to that of LLM heterogeneity.

Role–Specialization Alignment

Aligning social labels with agent specialization has minimal additional effect. While it marginally affects agent-level metrics such as plasticity and directedness, the consensus shift is actually dampened under role–specialization alignment relative to unaligned specialist populations.

Network Structure

Network structure (ER vs. WS) interacts with the effects of persona-style role prompting at the agent level: it modulates plasticity and directedness but leaves collective consensus unaffected. For scenarios involving LLM heterogeneity, the impact of network structure is statistically insignificant.

Modeling: Surrogate Opinion-Dynamics

A hierarchy of surrogate models, inspired by classical opinion-dynamics (DeGroot, Friedkin–Johnsen), is constructed to predict belief transitions and consensus formation:

• M1: Individual persistence,
• M2: Persistence + population-level belief composition,
• M3: Persistence + local neighborhood composition,
• M4: M3 + agent identity features (specialization and role).

Predictive accuracy, measured as Matthews Correlation Coefficient (MCC) on individual transitions, increases monotonically with feature richness—especially in scenarios involving specialists, where local neighborhood and identity information are relevant. Outcome-level fidelity, assessed as agreement between empirical and simulated consensus, is maximized by models that incorporate population-level belief composition. In homogeneous generalist populations, simple persistence-based surrogates suffice; in heterogeneous populations, dynamics converge only when collective belief-state features are present. This reveals a clear decoupling: the mechanisms that best explain agent-level transitions are not sufficient to reproduce population-level consensus in the absence of population/neighbor composition features.

Theoretical and Practical Implications

The results of this study indicate:

• LLM agent diversity, originating from genuine heterogeneity in underlying models (not merely persona labels), is necessary for realistic simulation of belief exchange and consensus dynamics in multi-agent LLM systems;
• Population-level phenomena cannot be inferred from agent-level update heuristics alone—collective outcomes result from the interaction of individual inductive biases, population composition, and local network effects;
• Classical models of opinion dynamics accurately capture diffusion in homogeneous LLM groups but are inadequate for heterogeneous agent ecosystems unless they incorporate higher-level belief composition features.

Practically, these insights have consequences for the design of multi-LLM platforms in autodebate, collaborative reasoning, and information infrastructure scenarios. Simulations constructed solely from prompt-diverse instances of the same model do not faithfully approximate the behavioral diversity or consensus dynamics of genuine heterogeneous populations. Future multi-agent LLM orchestration for applications in aggregation, deliberation, or content moderation must explicitly account for the epistemic priors and belief update dynamics induced by true model variance.

Limitations and Future Directions

The experimental grid is domain-specific (medical statements) and confined to Llama-3 and related variants with a moderate population size. Observational windows are limited to short interaction timelines, and belief states are inferred from output probabilities over a discrete set. The framework could be fruitfully extended to broader domains, variable network topologies, longer memory protocols, asynchronous messaging, and latent representation-level belief tracking. Comparative studies between LLM agents and human participants subjected to analogous protocols are also a natural extension.

Conclusion

This work provides a rigorous analysis of belief coevolution within social networks of LLM agents, systematically disentangling the respective effects of domain specialization, social roles, and network structure. The primary finding is that LLM heterogeneity is the critical driver of both agent-level change and collective consensus, while persona labels and network effects modulate the diffusion pathway but not the point of convergence. The results set empirical and theoretical baselines for future studies in multi-agent LLM dynamics and provide actionable guidance for the development and evaluation of AI systems involving interacting populations of LLMs (2607.27512).