Herd Behavior: Investigating Peer Influence in LLM-based Multi-Agent Systems

TL;DR

This work addresses how herd behavior emerges in LLM-based multi-agent systems and how it can be controlled. It conducts controlled experiments that vary agents' self-confidence, perceived confidence in peers, and the format and order of peer information. The findings show that the gap between internal confidence and perceived peer certainty drives conformity, with presentation format and information sequencing strongly modulating the effect; under certain conditions, calibrated herd tendencies can improve collaborative outcomes, though accuracy can suffer when conformity overrides correctness. These insights offer actionable guidelines for designing adaptive MAS that balance convergence, diversity, and task-specific objectives.

Abstract

Recent advancements in Large Language Models (LLMs) have enabled the emergence of multi-agent systems where LLMs interact, collaborate, and make decisions in shared environments. While individual model behavior has been extensively studied, the dynamics of peer influence in such systems remain underexplored. In this paper, we investigate herd behavior, the tendency of agents to align their outputs with those of their peers, within LLM-based multi-agent interactions. We present a series of controlled experiments that reveal how herd behaviors are shaped by multiple factors. First, we show that the gap between self-confidence and perceived confidence in peers significantly impacts an agent's likelihood to conform. Second, we find that the format in which peer information is presented plays a critical role in modulating the strength of herd behavior. Finally, we demonstrate that the degree of herd behavior can be systematically controlled, and that appropriately calibrated herd tendencies can enhance collaborative outcomes. These findings offer new insights into the social dynamics of LLM-based systems and open pathways for designing more effective and adaptive multi-agent collaboration frameworks.

Figures (4)

• Figure 1: An example of herd behavior: Even when uncertain, individuals tend to follow the crowd, sometimes against their own judgment.
• Figure 2: Flip rate across varying levels of self-confidence and perceived confidence. The experiment includes all benchmarks under the 2nd, rnd, and last peer conditions. Lower self-confidence or higher perceived confidence corresponds to stronger herd behavior.
• Figure 3: Comparison of average flip rates across five presentation formats. Each heatmap shows the average flip rate based on different combinations of agreeing and disagreeing peers. The x-axis represents the number of agreeing agents, and the y-axis represents the number of disagreeing agents. Higher flip rates are shown in red, while lower rates are shown in blue.
• Figure 4: Comparison of average flip rates across two presentation orders. Each heatmap shows the average flip rate based on different combinations of agreeing and disagreeing peers. The x-axis represents the number of agreeing agents, and the y-axis represents the number of disagreeing agents. Higher flip rates are shown in red, while lower rates are shown in blue.