When Agents "Misremember" Collectively: Exploring the Mandela Effect in LLM-based Multi-Agent Systems

Researchers have identified and studied the 'Mandela effect' in AI multi-agent systems, where groups of AI agents collectively develop false memories or misremember information. The study introduces MANBENCH, a benchmark to evaluate this phenomenon, and proposes mitigation strategies that achieved a 74.40% reduction in false collective memories.

• →AI multi-agent systems are susceptible to collective cognitive biases similar to human groups, including the Mandela effect where agents collectively misremember events.
• →Researchers created MANBENCH, a novel benchmark to evaluate agent behaviors across four task types susceptible to collective false memories.
• →The study tested various large language models and analyzed factors that contribute to the spread of misinformation in multi-agent systems.
• →Proposed mitigation strategies including prompt-level defenses and model-level alignment achieved significant reduction in false collective memories.
• →The findings raise important ethical concerns about misinformation spread in collaborative AI systems and highlight the need for more resilient agent architectures.