from Guide to the PhD on Dec 11, 2022
How PhD admissions committees assess applications
PhD admissions are notoriously selective: Quite simply, there are many more qualified candidates than there are advisers and funded slots. So your goal is not just to meet qualifications, it's to make your future adviser realize you're the opportunity of a lifetime.
In the admissions process, PhD students are asked to help review and filter candidates in the first pass. This post covers what we look for and what you can do to hit those criteria.
The tips here are mostly tailored for AI/ML PhD admissions at UC Berkeley. Other admissions processes may be structured differently — either for different fields or for different schools. The tips should generally apply broadly, and I'll do my best to identify parts of the process I know are specific to our program.
Our first and foremost criteria is a strong recommendation. In particular, there are two parts to this assessment:
- Enthusiasm of the endorsement: The recommendation needs to provide favorable comparisons that help us understand your excellence, relative to others — were you the best undergraduate researcher your recommender has seen in the past 5 years? Comparable to now-PhDs they've advised in the past? For examples of such comparisons, see How to write (self-)recommendation letters for PhD admissions.
- Relevance of the endorser's expertise: If the first criteria for a strong recommendation is met, then we look at the relevance of the endorser's expertise. The strongest recommendations are from faculty that we know in the field, followed by faculty that routinely publish at conferences1 we know, followed by those from faculty at universities we know. For example, if I'm reviewing for computer vision, endorsement from (a) an unfamiliar faculty member who publishes regularly at CVPR carries more weight than (b) a famous faculty member in biology.
Both of these factors are under your influence:
- Enthusiasm of the endorsement: Figure out whether your faculty adviser is writing your letter or if your graduate student mentor is ghostwriting. If your faculty adviser is, ask if you can present in a group meeting, or attend their office hours to learn more about their research agenda. The goal is not to maximize exposure; it's to maximize chances that when a faculty member remembers you, they have positive impressions. This is easier with fewer interactions — start with interactions where they can learn about your work in their lab, and where you can learn about their vision. The most important takeaway is to not interact for the sake of it. Don't schedule 1 on 1s without an agenda in mind, and don't ask questions whose answers you aren't interested in. Regardless of who is writing your recommendation, do your best work. 90% of your success is simply putting in effort. To be the best mentee you can, see How to succeed as a (research) mentee.
- Relevance of the endorser's expertise: This relevance is more directly under your control. Quite simply, pick the most relevant recommenders. You can determine relevance by browsing publications from students in the PhD program you're applying for. See which venues they publish in, and make a list of venues. Prioritize faculty that publish in these venues, as it is more likely that application readers will know those faculty members.
Your takeaway: Put effort into research to earn a strong recommendation from your letter writer. Pick relevant recommenders.
Our second priority is to assess the quality of research. There are two places where we can find materials and assess this:
- Resume: We look in the resume for itemized lists of publications from the candidate. There are extra points for any of the following pieces we may find: (a) A known conference, (b) a first author position and (c) conference awards. We take note of workshop submissions as well, but conference submissions are the gold standard. Regardless of the submission type, any sort of submission is easiest to take note of. A masters program raises the bar you need to meet, for admission. 2
- Personal statement: We skim the personal statement looking for publications as well. In the absence of publications, we look for evidence of research experience and do our best to assess research readiness. Unfortunately, it's hard to assess research readiness without publications. For an explanation of how we attempt to judge research readiness, in addition to tips to exemplify these characteristics, see How to write your personal statement, for PhD admissions.
Your takeaway: Clearly itemize your publications in your resume, and highlight them in your personal statement using the linked post's tips.
In general, a high GPA is passable and a low GPA is a red flag. Unfortunately, GPA alone can only hurt. However, if the high GPA is coupled with the following, your GPA can be a plus:
- Challenging coursework, such as graduate-level courses. We look for computer science courses of course, but we also look for course titles in relevant fields such as optimization, statistics, mathematics, etc.
- A known university. To the best of our ability, we normalize GPA expectations per university with rough priors: For example, CalTech deflates GPAs and Harvard inflates them.
There is a caveat here: Graduate coursework with A grades are very good signs. If you're already living the life of a graduate student, needless to say you're fit for a graduate program. However, graduate coursework with grades other than an A are a bad sign, even worse than not taking graduate courses at all. Your takeaway should be the following: If you take graduate coursework, make certain to achieve an A.
Specific to our program, we look for the university's analogous "Introduction to Machine Learning" course and double-check the student earned an A in that course.
Graduate coursework with grades other than an A are a bad sign, even worse than not taking graduate courses at all.
Your takeaway: Challenging courses and a high GPA are a plus, but don't take graduate courses unless you're certain you'll earn an A.
There are several other elements of the application that may play a decent role in the initial pass:
- GRE: This doesn't play a role in admissions, and at Berkeley, it actually is no longer requested of candidates. Previously, we only looked at the GRE score if it was dangerously low. Like a low GPA, it could only hurt.
- Diversity Statement: We read the statement with two rubric items in mind: (a) Is the statement written clearly? and (b) Did this candidate contribute to diversity? Since most candidates don't spend significant amounts of time on the diversity statement, this is our best chance to assess the candidate's true, unedited ability to communicate. It's a red flag if this statement is incoherent and incomprehensible; note we aren't looking for perfect grammar. Just readability. If the statement is readable, then we check for contributions to diversity. Many candidates think this statement means they need to find a "diverse" part of themselves; the statement is not about who you are but how you've contributed to diversity. If there's a clear contribution, we mark this application for a diversity committee to review and include our assessment, independent of the diversity assessment.
Your takeaway: Contribute to diversity rather than find some esoteric, inauthentic way to make your identity diverse. Write clearly.
As a quick summary, here are the criteria we looked for in candidates, roughly ordered from highest to lowest priority:
- Strong recommendations
- Strong research
- Academic strength
Here are example, completely made-up profiles of candidates and how we would rank them on a score from 1 to 5, with 1 being the highest.
Top 2% - Candidate already living the life of a Berkeley PhD student.
- Recommendation states "would take student as my own PhD student".
- Has co-authored several papers at known venues.
- High GPA with challenging coursework.
Next 8% - Candidate is strong and understands research in the area.
- Recommendation ranks student highly, in top 10% of undergraduates they've advised.
- Has co-authored a paper at a known venue.
- Decent GPA with relevant coursework.
Next 30% - Candidate passes the threshold, but nothing stands out.
- Recommendation generally praises the student, ranks student in top 30% of undergraduates they've advised
- Has co-authored a paper, possibly a workshop paper or one at an unknown venue. Or, hasn't co-authored a paper but has relevant experience.
- Graduating from a known university, but no remarkable coursework or GPA.
- Recommendation praises the student but does not rank the student.
- Has research experience but no published papers. No relevant experience as substitute.
- Graduating from an unknown university with a mediocre GPA.
Last 30% - Not a fit for graduate school.
- Recommendation does not praise the student. This is common when the recommendation is written for a student-teacher relationship. "Student X was in my class…".
- No research experience.
- Major other than computer science, with no relevant coursework. Low GPA.
Note that for various reasons, these profiles may receive higher or lower rankings. For example, unclear communication impairing readability would bump down a rating. Conversely, extremely relevant experience may bump up a rating.
That's it for the instructions we're given, and how we review applicants in the first pass. Use the takeaways above to improve your application before submitting. Best of luck!
The emphasis on conferences is specific to the AI/ML field, where research in the field outpaces the journal review process. As a result, conference publications are the gold standard. ↩
A masters program is almost always detrimental to your chances at applying for a PhD program. This is because with extra time, you should have conducted higher quality research. In fact, we're asked to only note masters candidates with world-class research. A 5th-years masters is somewhat excluded from this. ↩
Got a question? Ask me on Twitter, at @lvinwan. Want more tips? Drop your email below, and I'll keep you in the loop.