Different models can make different errors or offer complementary strengths. A combination rule determines how their contributions affect the final output; the models need not communicate as autonomous agents.
For example, several models can draft answers that a separate stage ranks and combines. Additional calls increase resource use, and a weak judging step can select errors or hide disagreements.
ELI5
Imagine asking several helpers the same question and then deciding how to use their answers together.
For example, one helper may catch a missing requirement while another supplies a clearer explanation. Agreement alone still does not prove an answer is correct.
Must an ensemble use voting?
No. It can vote, average predictions, rank candidates or generate a synthesis, depending on the task.
Does combining models always improve quality?
No. Correlated mistakes, poor selection and extra cost can outweigh the benefit of multiple contributions.
