I'll comment on RT/resolution scale as that's where the more immediate difference can be noticed aside from how blurry 1440p vs any DLSS 4K setting is.
First up is the various RT Lighting settings. Keeping it simple; RT lighting medium<RT lighting Ultra<RT lighting Psycho. The higher your setting is for this parameter, the more lighting data the API has to work with. That means noise is reduced because it's not having to interpret the difference between shadow/light as much and that's where noise comes from. It's why it is most noticeable in the evenings, alleyways, under bridges, etc. The major benefit to running Psycho is denoising quality because the differences otherwise are quite minuscule - GI bounce lighting is not that noticeably different over standard lighting from what I've seen because it doesn't override the game's base faux GI data. (There was a noise issue introduced in 1.5 with skin shaders caused by the RT radius adjustments the devs made. It's noteworthy, but not entirely what I'm getting at here.)
Semi-gloss, high-frequency texture, reflection noise is a different beast that seemingly needs more development on the API side - it's an issue in the few games that have the effect implemented and it's there even at 4K native. Perhaps it's not at 8k but I don't have an 8k monitor, nor do I have a 3090Ti, so testing it isn't even an option for me. Take a look at the patch work on the road in the second image, left of center. Larger, fuzzier, reflection splotches at balanced than quality. Is the game ruined? No. Did you even notice? Doesn't seem like it. It's an example of RT quality effected by resolution and scaling though. Even though it's a "reflection", the texture density coupled with high contrast between shadow/light create a scenario that noise can't be avoided - The API simply can't understand what you're looking at in the given resolution. We can get more into why it's here and not in actual reflections, such as water or clear coats another time perhaps.
Back to lighting specifically - RT lighting quality is impacted by resolution as well. Assuming native resolution, just to keep it simple; RT Ultra at 4k should present roughly the same, perhaps even less, noise as RT Psycho at 1440. RT Psycho at any resolution will render better than any of the other RT settings at the same scale just as RT Psycho at quality DLSS will render better than RT Psycho at balanced and it will render best at native. I can see better diffuse AO and local diffuse lighting in your images at quality vs balanced. You didn't seem to notice it though, so does it really matter?
It's a similar case for volumetric fog. Resolution - DLSS quality x volumetric fog quality, or something to that effect, determines how prevalent aliasing of emissive volume data will be.
I can't go over all of these details with images as I don't have the game installed at the moment.
Seeing as you're running RT Psycho lighting, meaning your have some frames to spare, I'll make a suggestion for you.
Set RT lighting to Ultra, or keep Psycho if you really insist, and create a new ini file in the engine\config\platform\pc folder.
Inside the ini type the following:
[RayTracing]
EnableImportanceSampling = false
Ever look up at an overpass in the middle of downtown in the mid night (Near Embers comes to mind) and see the diffuse lighting abruptly cutoff to start darkness? Did you walk forward and notice that more diffuse lighting was filled in with your movement? That's importance sampling at work. The game's obnoxious volumetric fog hides this to a certain degree, but it will be a requirement for my Ghost In The Shell env mod since the fog is reduced to a level where it can't be unnoticed with RT lighting on at any setting.
