🤖 More Agents Is All You Need

The paper reports that simple sampling-and-voting across multiple agents improves large language model performance, and that this performance scales with the number of agents instantiated. The method, called Agent Forest, is orthogonal to existing more complicated methods for enhancing LLMs, and the size of the gain is correlated with task difficulty. The authors, Junyou Li, Qin Zhang, Yangbin Yu, Qiang Fu and Deheng Ye, say they ran comprehensive experiments on a wide range of LLM benchmarks to verify this finding and study the properties that facilitate it. Their code is publicly available.