Discussion about LISA

In the article, only the comparison of the average weight paradigm of each layer during lora fine-tuning is given.

1. But what if their weights are different before fine-tuning?
2. Using the weighted averaging method, is it possible that there are differences in the intermediate iteration process?