Questions about LOD bias and texture streaming pool

Hi guys, newbie here, really need help from staff and experienced ones.
I am having problems about memory optimazition in my project, where I need to resize some textures in one level while using the original ones in another. For example, I need to use 10241024 (original) textures in the first level but need them in 256256 in the second level. To my knowledge, I should operate mips data to get what I want.

I noticed there seems to be two kinds of LOD bias for textures and materials. One of them is calculated using LOD_group and LOD_bias in texture properties, when you cook and package your game. And this cannot be changed at runtime. Another is set in material editor->texture sampler, which accepts dynamic input at runtime.

So here are my questions:
Am I right about this “two kinds of LOD bias” thing? Can I set LOD bias in texture samplers AT RUNTIME and get any memory saving?
If I am wrong, is there any tricks to avoid loading the first several mips into the pool, or this is not recommended and supported by unreal?