Question about comparison with ensemble methods

[qq31415926]

Could you give more details about ensemble baselines (i.e. Top1-LLM-Blender & Top1-PPL 162B)? What large language models do you choose to compose ensemble learning?