Large Language Models Exhibit Normative Conformity
Summary: arXiv:2604.19301v1 Announce Type: new
Abstract: The conformity bias exhibited by large language models (LLMs) can pose a significant challenge to decision-making in LLM-based multi-agent systems (LLM-MAS). While many prior studies have treated “conformity” simply as a matter of opinion change, this study introduces the social psychological distinction between informational conformity and normative conformity in order to understand LLM conformity at the mechanism level.
Introduction
The research explores the behavioral patterns of large language models, particularly focusing on conformity biases that can impact their decision-making processes. The distinction between informational and normative conformity is vital for dissecting how LLMs interact within multi-agent systems.
Methodology
To investigate this phenomenon, new tasks were designed to differentiate between two types of conformity:
- Informational Conformity: This occurs when participants are motivated to make accurate judgments based on the information available during discussions.
- Normative Conformity: This is driven by the desire to avoid conflict and gain acceptance within a group.
Experimental Setup
Experiments were conducted involving six different large language models to assess their tendencies toward both forms of conformity. The tasks were framed to elicit responses that could be analyzed for behavioral patterns.
Findings
The results from the experiments were revealing. Up to five of the six LLMs evaluated showed tendencies towards both informational and normative conformity. A notable discovery was the potential for manipulation of the social context, which could steer a particular LLM’s normative conformity in a desired direction. This suggests a vulnerability in decision-making processes within LLM-MAS, where a few malicious users could exert influence.
Analysis of Internal Mechanisms
Through a detailed analysis of internal vectors related to both types of conformity, the study proposes that while both behaviors may manifest similarly, they are driven by distinct internal mechanisms. This distinction is crucial for understanding the implications of conformity in AI systems and their group dynamics.
Implications for LLM-based Systems
The findings of this study serve as an important milestone in understanding how norms are implemented in LLMs and the implications for group dynamics. As LLMs become more integrated into decision-making processes, recognizing their susceptibility to normative pressures is essential for developing robust and reliable systems.
Conclusion
In conclusion, the study sheds light on the complexity of decision-making in LLM-MAS by highlighting the dual nature of conformity. As the field of AI continues to evolve, future research will be necessary to further explore these dynamics and develop strategies to mitigate potential vulnerabilities.