It is basically the requirements for getting a job as a scientist in my field (machine learning) as far as graduating goes:
1) 1 journal paper as the first author, so that they learn how that process (often is a review or perspective paper on their field, which goes in their dissertation)
2) 2 first author papers in top-tier machine learning venues (goes in dissertation)
3) 1 first author paper in a top-tier or second tier venue (goes in dissertation)
4) 1 collaborative paper with another PhD student (have to learn how to collaborate)
By the end, they have at a minimum around 4 first author papers and 1 additional paper. They can then turn these into their dissertation and these are signals to employers that they are competent scientists. If a student tells me they want to be a faculty member, we increase the numbers a bit (need at least 10 first-author papers to be competitive).
I also require that they be organized when we have our one-on-ones in terms of their experimental output.
That's pretty much it. It may sound like a lot, but I try to put training wheels on for the first couple projects and make them as tractable as possible because I fleshed out the project and found some low-hanging fruit, before I start trying to make them be more independent and drive the process.
That's an insane workload to put out during a PhD. 5 Papers including 2 top-tier ones in 3-4 years? Is ML really full of such low-hanging fruit where that is possible?
In my stints around Denmark and Japan (Computational Chemistry) you basically need ~1 top-tier or ~3 mediocre papers during your 3-year PhD.
Norms vary per academic discipline and niche. My requirements are pretty much the standard for my field if you want a job doing research at FAIR, DeepMind, Adobe Research, Google Research, OpenAI, etc. Without meeting those goals students will not be able to get research oriented jobs at top places.
In the US, PhD students in Computer Science and similar fields usually require 5-7 years to finish. They typically do not have MS degrees, unlike in Europe, where the expected graduation time is 3-4 years and they must have MS degrees.
You would be surprised how often there is a lot of low-hanging fruit. I'm good at asking "easy" questions that nobody knows the answer to, so a student just has to put in the work to get the answer. Many advisors don't give as much initial hand holding as I do.
So far my PhD students have taken 3-6 years to finish. Those that have graduated so far have finished with 3 papers (3 years), 9 papers (3 years), and 12 papers (5.5 years). The senior ones still in my lab (year 4 or 5) are on track to have around 3 first author papers in top-venues and about 9 papers total each. Their first author papers are in CVPR, ICCV, ACL, ECCV, BMVC, NAACL, ICLR, AAAI, etc.
Thanks! That makes sense. My requirements for my students aren't quite at the same level. (And I arguably didn't meet those myself for my own PhD.) 10 first author papers is seems a little crazy to me. As a relatively new faculty member, I haven't met that bar myself yet. But of course as you say in another comment below, it differs by field.
