The fear of an incremental label is real, but it is also a symptom of a deeper misunderstanding about how research is evaluated. Reviewers do not penalize a paper for being a piece of a larger whole; they penalize it for being a thin slice that offers no standalone value. Your concern about the ArXiV preprint is not unfounded, but it is misdirected. The risk is not that you split the paper or that you changed the title. The risk is that you submit a fragment that reads like a fragment, leaving reviewers to do the work of connecting dots you never drew for them.
Here is what this means in practice: your framing is not a cosmetic choice, it is the substance of your submission. When you split a paper, you are not dividing a dataset or a proof into two equal halves. You are creating two distinct contributions, each of which must answer a basic question on its own: what did you learn here, and why should anyone care now? If you cannot answer that in the abstract, in the introduction, and in the final discussion, then reviewers will sense it, and they will call it incremental, not because your work is small, but because you presented it without its own spine. The ArXiV version is not your enemy. It is actually a tool. Reviewers can find it, yes, but they can also see that you are not hiding it. The problem emerges when you pretend it does not exist, because then you force them to discover it on their own, and that discovery feels like a gotcha rather than a natural part of your research arc.
So what do you do? You stop treating the preprint as a liability and start treating it as a reference point for clarity. In your NeurIPS submission, you can acknowledge the broader work without undermining your current contribution. You can say, in effect, "This paper focuses on X, while the full line of inquiry extends to Y." That is not a confession of weakness; it is a demonstration of intentionality. Reviewers respect a researcher who knows exactly where the boundaries of a contribution lie. What they do not respect is a submission that tries to erase its own history, because that erasure is transparent and it signals that you are more concerned with perception than with progress.
The practical takeaway is this: do not let the fear of the word "incremental" push you into a defensive posture. Instead, use the split as an opportunity to sharpen your claims. Write the introduction as if your readers have never seen the ArXiV paper, because they probably have not, and then write the conclusion as if they just read your NeurIPS draft, because that is the only thing they will evaluate. If you can make the first half of your work feel like a complete meal rather than a leftover, you will have answered your own question. Yes, the doubt is reasonable. But it is also surmountable, provided you treat your own paper as the primary artifact of your research, not as a placeholder for something bigger.