More Agents Is All You Need

Researchers have discovered that the efficiency of large language models (LLMs) can be improved through a straightforward sampling-and-voting method, which scales performance in correlation with the quantity of agents involved. This technique offers a simplication compared to existing complex methods, potentially enhancing LLMs even further depending on the task's complexity. Conducted experiments across various benchmarks confirm these insights, with the code made available for public access.

Visit Original Article →