Empathy

Concepts:
There are neurons called 'mirror neurons': they help you empathise with other people, as well as understand their actions. They work, broadly speaking, by simulating, in a rudimentary way, the thought process behind what you are trying to understand. Mirror neurons for an action fire in a way that mimics how, say, other neurons fire as you perform that action, when you see someone else doing the same action. Likewise, when you see someone, say, being hurt, a corresponding mirror network roughly processes how they would respond to it - the results can be felt by you, as you may wince when someone else is hit, or feel bad when something bad is happening to them.

Creation:
So how does this relate to tulpas? To put is simply, a tulpa is a complex mirror network.
You come up with a personality for a tulpa, and think about what they would do in certain situations - in other words, you are consciously performing the job of a mirror network. We all know that this is how we learn things: you start off consciously, deliberately doing it, and over time it becomes 'hardwired' - sometimes motor memory, here in a mirror network. This process is repeated when you parrot: it is obvious that you think about how your tulpa would respond (bad therefore how they would think).

Narration; talking to your tulpa. Crucial for most, but how does it fit in? Conversation. While having a conversation, do you not often think about how someone would respond to something that you say? This can be done either consciously, or unconsciously - via mirror networks. In talking to your tulpa and treating it like a conversation, you are doing the same thing. Consciously or not, you are stimulating your tulpa's mirror network, which stimulates development correspondingly.

Spending time with your tulpa in general does the same thing. You think about what they might do, which is another, basically the same, decision-making process as what they would say or feel, and therefore also relies on mirror networks. It is clear that the same things apply to movement as have been said about speech above.

The more you use a network, the stronger and more complex it becomes. This is true for mirror networks - instinctively (i.e. with mirror networks) it is much easier to predict someone who you have known for a long time. Like it being easier to speak a language, this means that:
1) This network has been used a lot
2) It is correspondingly well-developed

Of course, once the network grows enough it is capable of complex thought in itself. For as long as you think about your tulpa, the network is stimulated because you expect something to happen but do not know what. This helps to explain why many tulpas find it difficult to interject when their host is not thinking about them, as well as why early on, tulpas can feel predictable and hardly unconscious - because that's exactly what they are.
It also provides a preliminary entry point for the role of expectation: the more you expect your tulpa to respond (i.e. the more you believe they are sentient) the more you predict the response, so the more the network is stimulated.

Evidence:
Many early responses are empathetic in nature. Specifically, I want to draw attention to the emotional response. You feel a wave of foreign emotion; it is not 'yours' - i.e. it was not generated by your personal processing of your experience - and yet it is clearly within the mind. What this sounds like, to me, is a strong empathetic response, a mirror network grown stronger than we would normally see.
Another phenomenon: has your tulpa ever experienced a strong emotion and you reacted physically to it? Even without feeling sad, I have involuntarily cried when my tulpa did, and I know that I am not alone. Again, this seems (especially given the fact that this, like an emotional response, is not generated by the host's personality) to be the result of empathetic machinery.

Conclusion:
And what does this mean? It means that personality is more useful than previously thought; it means that thinking for your tulpa is definitely not a bad thing. It also lays the groundwork for tulpas to be, in the end, neurologically equal to the host - since a sufficiently advanced imitation is as good as the real thing, so is a sufficiently complex network for simulating someone else's thought essentially someone else's mind.