When agents collectively "misremember": Exploring the Mandela effect in LLM-based multi-agent systems

The researchers systematically study the Mandela effect in LLM-based multi-agent systems, namely the phenomenon in which a group of agents collectively misremember past events under social influence and internalized misinformation. The team proposes the MANBENCH benchmark, covering four categories of tasks susceptible to this effect and five interaction protocols that vary agent roles and memory timescales, and quantifies this effect and analyzes influencing factors on multiple LLM-driven agents. The paper also proposes prompt-level defenses such as cognitive anchoring and source scrutiny and alignment-based model-level defenses, reducing the Mandela effect by an average of 74.40% compared with baselines. Code and dataset are available at https://github.com/bluedream02/Mandela-Effect。