reason of averaging model

[angus5552]

Hello

I am training Cosyvoice2.0 model
during training I have question regarding training

In the run.sh, after training each model, script trying to average all models
I tried to search reason on the paper, but could not search the reason.
(researched Cosyvoice2.0 and Cosyvoice 1)

Could anyone tell me why trying averaging model after finishing training model?

[aluminumbox]

this can increase model performance normally

[angus5552]

Thanks for response :)
Could you share me any papers or article regarding effect of averaging models?