Are modern ML PhDs becoming too incremental, or is this just what research looks like now? [D]

I’ve been thinking about the current state of machine learning PhDs, including my own work, and I’d like to hear how others see it.
My impression is that a large fraction of modern ML PhD work follows a fairly predictable pattern: take an existing idea, connect it to another existing idea, apply it in a slightly different setting or community, tune the system carefully, add some benchmark results, and present the method as a new state-of-the-art approach. Another common pattern is mostly empirical: run benchmarks, report observations, provide some analysis, and frame that as the main contribution.
To be clear, I’m not saying this work is useless. Incremental progress matters, and not every PhD needs to invent a new paradigm. But sometimes it feels like many ML PhDs are closer to extended master’s theses: more experiments, more compute, more polished writing, and more benchmarks, but not necessarily a deeper scientific contribution.
What bothers me is that the same pattern appears even in top-tier conference papers. A paper may look strong because it has a clean story, a benchmark win, and good presentation, but after removing the “SOTA” claim, it is not always clear what lasting knowledge remains. Did we learn something general? Did we understand a mechanism better? Did we identify a failure mode? Did we create a reusable method or evaluation protocol? Or did we mostly produce another temporary leaderboard improvement?
I’m also reflecting this back onto my own PhD. I see some of the same patterns in my work, so this is not meant as an attack on others. It is more of a concern about the incentives of the field. ML seems to reward publishable deltas: small method variations, new combinations, benchmark improvements, and convincing empirical stories. But I’m less sure whether it consistently rewards deeper understanding.
So my question is:
Have ML PhDs become lower-quality compared to PhDs in other fields, or is this simply the normal shape of cumulative research in a fast-moving empirical field?
And maybe more importantly:
What separates a genuinely strong incremental ML PhD from one that is basically a collection of polished benchmark papers?