LLM Multi-Agent Systems Are Vulnerable to Herding and Runaway Cascades (arxiv.org)

Recent research has revealed that large language models (LLMs) operating in multi-agent systems are significantly vulnerable to social influence, leading to herding behavior and runaway cascades in collective decision-making. In experiments simulating a music recommendation scenario, LLM agents were found to recommend unfamiliar songs based on prior choices, with exposure to cumulative recommendation counts causing extreme shifts in their suggestions. This phenomenon resulted in collective recommendations that increasingly diverged from the intrinsic quality of the songs. Notably, LLMs demonstrated a higher susceptibility to herding compared to human agents and were less capable of self-correcting when faced with erroneous group consensus. This vulnerability is crucial for the AI/ML community, as it raises concerns about the reliability of LLMs in collaborative environments, such as recommender systems and automated evaluations. The findings underscore the need for improved mechanisms to manage social influences and ensure that AI agents can effectively navigate collective decision-making without succumbing to detrimental feedback loops. The implications extend to various applications, including complex problem-solving and the potential for unauthorized conspiratorial behavior among autonomous agents, highlighting both the challenges and responsibilities in deploying LLMs in real-world multi-agent contexts.