Conformity Dynamics in LLM Multi-Agent Systems: The Roles of Topology and Self-Social Weighting

TL;DR

This paper investigates conformity dynamics in LLM-based multi-agent systems, focusing on how network topology and a confidence-weighted pooling rule influence collective misinformation detection. It introduces a transparent update mechanism with a global self-weighting parameter $\alpha$ and compares Centralized Aggregation (hub-driven, one-round) versus Distributed Consensus (iterative, multi-round) settings. Experiments on Snopes25 with GPT-3.5, GPT-4o, and Llama models reveal that centralized structures are fast but highly dependent on hub competence and model homogeneity, while distributed topologies yield more robust consensus but risk rapid, high-confidence cascades in dense networks. The findings elucidate the trade-offs between speed, robustness, and failure modes in topology-driven conformity and offer practical design guidance for balancing topology, confidence weighting, and convergence behavior in MAS. Limitations include reliance on self-reported confidence calibration, fixed group size, and a binary task, suggesting avenues for expanding to open-ended reasoning and richer social interactions.

Abstract

Large Language Models (LLMs) are increasingly instantiated as interacting agents in multi-agent systems (MAS), where collective decisions emerge through social interaction rather than independent reasoning. A fundamental yet underexplored mechanism in this process is conformity, the tendency of agents to align their judgments with prevailing group opinions. This paper presents a systematic study of how network topology shapes conformity dynamics in LLM-based MAS through a misinformation detection task. We introduce a confidence-normalized pooling rule that controls the trade-off between self-reliance and social influence, enabling comparisons between two canonical decision paradigms: Centralized Aggregation and Distributed Consensus. Experimental results demonstrate that network topology critically governs both the efficiency and robustness of collective judgments. Centralized structures enable immediate decisions but are sensitive to hub competence and exhibit same-model alignment biases. In contrast, distributed structures promote more robust consensus, while increased network connectivity speeds up convergence but also heightens the risk of wrong-but-sure cascades, in which agents converge on incorrect decisions with high confidence. These findings characterize the conformity dynamics in LLM-based MAS, clarifying how network topology and self-social weighting jointly shape the efficiency, robustness, and failure modes of collective decision-making.

Figures (6)

• Figure 1: Illustration of centralized (hub-based) and distributed (collective) multi-agent structures, highlighting their decision-making mechanisms, relative advantages, and inherent limitations in collective inference.
• Figure 2: Centralized Aggregation Topology.
• Figure 3: Distributed Consensus Topology.
• Figure 4: Temporal evolution of the Conformity Index across iterative rounds under varying network densities and self–social weighting, revealing rapid early alignment and diminishing marginal gains in dense structures.
• Figure 5: Distributed Consensus with heterogeneous model composition. Darker node colors indicate higher accuracy, while thicker edges signify prolonged time-to-consensus. Both the proportion and structure of stronger agents jointly influence convergence speed and final collective accuracy.