This part of the brain is starting to get some recognition as basically what makes us human. It might be helpful in answering this question from a neurological standpoint. If anyone knew how it worked. Which no one does. But maybe some day!

If you're using the "self-generated responses versus expected behavior" groups, I'm really of favor that a tulpa is a combination of both, their responses switching between the two methods constantly, depending on what their response/action actually has to be. I think tulpas are primarily powered by self-generated responses that we either A. consciously create and then lie to ourselves that the tulpa is saying it, B. take a response that we were originally going to say, but unconsciously rerouted the response to the tulpa, or C. most likely a combination of both. I think if we really have to think of what the tulpa's response will be, then we'll actually have to consciously think of it, but we also might find the response so intuitive or natural that the tulpa will "automagically" say it without us having to consciously think about it (which would actually fall under "expected behavior"). Ultimately, their replies are thoughts we were ultimately going to have, but we trick ourselves into thinking the tulpa is giving them. It's not so much of a mystical and vague "tulpa are frontends to our unconscious thoughts", but more directly "tulpa are our conscious/unconscious hyrbid puppets".

Another aspect that comes into play with tulpas is what people call Rubber Duck Debugging. By talking with a tulpa, you painstakingly detail every aspect of a given subject as opposed to working on countless assumptions by thinking about a subject in your head by yourself, assuring you don't skip over any critical data you assumed was meaningless. This, in combination with both self-generated responses and expected behavior, has the potential to make an extremely realistic alter-ego aka tulpa.

Of course, the ratio of self-generated responses versus expected behavior and the general immersion of how realistic the tulpa can be varies from person to person and depends on the person's current mental state. I'd imagine if you're more relaxed or otherwise closer to being in a suggestive state than not, then the expected responses would be more common, kind of like how dreams are nothing but expected responses. But ultimately, tulpas are probably a combination of methods rather than a single one.

I know all of my most special "holy shit this thing can act on it's own" experiences can be traced back to the rubber duck debugging theory. To this day, mine surprises me with information about myself that even I didn't know, but really, I'm probably just thinking really hard, harder than I would by myself, about aspects of myself, and then, unconsciously or not, relaying that information from my unconscious (aka assumed information that I haven't yet validated, otherwise known as a thought in mid-formation) to my consciousness via my tulpa (aka saying the information to myself to find out it sounds right). Mine has also actually helped me with bugs/features in my code, again, probably the rubber duck thing.