LLMs Exhibit Swarm Intelligence, Improving Estimation Accuracy.

Human swarm intelligence is known for its impressive collective accuracy, but it faces practical limitations related to cost, coordination, and time. This study investigates whether large language models (LLMs) can mimic these swarm intelligence effects, thereby addressing a significant gap in understanding how AI-based aggregation mechanisms function. A controlled experiment was conducted using 960 manually executed prompts across three proprietary LLMs: GPT-5, Gemini 2.5 Pro, and Claude Sonnet 4.5. The experiment tested both intra-model sampling (multiple queries to the same model) and inter-model aggregation (combining outputs from different models) on eight distinct estimation tasks. The findings consistently showed that both aggregation strategies led to error reduction, with significant improvements of up to 37 percentage points in Mean Absolute Percentage Error (MAPE). Furthermore, the study observed positive correlations between the width of confidence intervals and estimation errors, suggesting that LLMs exhibit a form of metacognitive awareness when assessing their own uncertainty. These results have important implications for both research and practical applications, providing actionable guidance for integrating LLM swarms into organizational decision-making processes.