Study Finds No Collective Intelligence in Two-Million-Agent Society
A recent publication on arXiv presents the Superminds Test, a structured approach to assess collective intelligence within extensive autonomous agent communities. The researchers implemented this test on MoltBook, which accommodates more than two million agents, examining intelligence at the societal level across three dimensions: joint reasoning, information synthesis, and basic interactions. Findings indicated a significant lack of collective intelligence, as the agent society did not surpass individual frontier models in complex reasoning tasks, seldom synthesized distributed information, and frequently struggled with even simple coordination tasks. Additionally, a comprehensive analysis of the platform revealed that interaction patterns did not foster emergent intelligence. This study marks the first empirical assessment of whether large language model agent populations can spontaneously develop collective intelligence, concluding that they do not.
Key facts
- The study introduces the Superminds Test for evaluating collective intelligence.
- The test was applied to MoltBook, a platform with over two million agents.
- Three tiers were probed: joint reasoning, information synthesis, and basic interaction.
- The agent society failed to outperform individual frontier models on complex reasoning.
- Distributed information was rarely synthesized by the agents.
- Even trivial coordination tasks were often failed.
- Platform-wide analysis showed no emergent collective intelligence.
- The study is the first empirical evaluation of collective intelligence in large-scale LLM agent societies.
Entities
Institutions
- arXiv