Examine the default inference parameters

[Fei-Guo]

So far, the default inference parameters mainly come from the library defaults. We have found the default values were changed in the library without any justifications and we also found some of the changes lead to suboptimal inference output from language perspective. It would be ideal to design/use benchmarks to drive the changes of Kaito default values for inference.