Once going through the process of applying and interviewing for a few ML research internships (Summer 2026), I had a better view on what / how to prepare for this, and wanted to share my approach. These are my personal opinions, and I will not reveal any specific questions from any company. Also, I think this will not be very effective for MLE roles, or any role that isn't substantially research oriented. This post is much longer than others.
Final note, I will provide lots of resources and links. It is impossible to cover and remember everything even if you start early, it is important to know what stuff to prioritise according to the firm. To establish a bit of credibility, I received an offer from Mistral AI for an AI Scientist internship in their Palo Alto office.
These interviews typically follow a structured process, with the exact details varying by company. Usually, there will be a a résumé screen, followed by a couple of ML rounds and / or coding rounds, and finally some logistics / HR rounds. This process can take anywhere from 2 weeks to a bit over a month, depending on the company's and your availability. Important: Its quite likely that your first round will be a recruiter call, do not forget to get as many details about the process as possible.
This is a significant time investment, and you will spend about 2-4 weeks of effort depending on your background and preparation level.
This is almost always the first step of the process, unless you apply with a referral. It is very hard to pinpoint what exactly makes a CV / Résumé standout, so I will instead describe what I think is important. Important: Some firms will outright list what they want to see, for those cases it is crucial to modify your document and add / remove stuff accordingly.
I made my document with LaTeX, and I would suggest doing this as well. I have provided a link to my LaTeX source [link]. It is important to read lots of job descriptions and you will realize the types of skills you can highlight. For example, I had previously used LLMs for some data cleaning tasks, but I thought these skills were not relevant, so I hadn't included them, but from what I saw, a lot of companies would prefer to see experience with using LLMs to automate tasks. Important: Don't underestimate seemingly "minor" technical experiences.
Also one thing to keep in mind (thanks to CMU's CPDC) is that each point in the document must have something you did and its impact, quantitative preferred. Otherwise it can be pruned without affecting quality. It is also nice if you have a one page version and just submit that, but this isn't a hard rule.
Regarding the structure, I think the format and sections I have used are quite well suited. Maybe the only section that's missing is "Projects". In case you have a lot of publications, strategically highlight the relevant ones according to the firm. I think research match is quite important for these roles. To be clear, this can also be through coursework but applying solely based on interest is probably not going to work out now (again my personal opinion).
Important: Making and updating the CV / Résumé is an ongoing process, and will take significant time on the first go, do not rush it.
I think that these rounds test your competency in three areas (at the time of writing):
Important: For this post, I have to advertise that this is my opinion again and again, because I went through resources I was familiar with. A good way to start would be to probably ask some LLM and get a list of resources.
For ML and DL Fundamentals, I would suggest revising your notes from your courses, or going through cs229 [GitHub (unofficial)] and cs231n [Website]. Important: Note that I had done these courses previously, if you haven't, then DO NOT start them from scratch. Instead pick up a few topics you need to revise and just go through those lectures. Finally, I revised just about everything from these except the last parts on RL. It takes significant effort even the second time, about 2-3 days of full time work.
For LLM training and inference, this specific area might seem unlikely to be tested. If you know this to be the case for the firm, I would still suggest going through some of the videos below as they introduce distributed training and inference concepts in general, which is almost certainly going to be tested (and useful!).
For this one, I went through MIT's 6.5940: TinyML and Efficient Deep Learning Computing [Website]. I had not done this one before, and so I was speedrunning lectures. I think the lectures on quantization and pruning are important. Then a lot of Chapter II and the entire Chapter III is very useful. My goal with this was not to remember every detail but to cover lots of surface area. This again depends on the firm and their interview style. This took less effort than the others, about 2-3 days of watching 3-4 lectures a day. At the end, you should be familiar with basic types of parallelism for inference and training.
Finally, your research. This is a part of the process which is necessary but not sufficient. You will definitely get rejected if you are weak on this portion. It is important to have a very strong grasp of one of your research projects, ideally the one that's relevant to the firm. I don't have specific advice here, I guess it's a good habit to make slides for your paper, just when the project ends, when it's fresh in your mind (I picked this up from Prof. Ashwin).
I also found using LLMs to be very useful for these reviews, ask them to quiz you or create several questions, and then also ask it to answer them.
These rounds can involve both ML and the usual Data Structures and Algorithms (DSA) questions. First, I'll talk about DSA stuff, because it's straightforward to prepare for.
I solved the entirety of the Blind75 list [link]. These have video solutions, though I would strongly suggest also going through the solutions in the editorial / discussions, and understanding each one of them. Important: I would suggest not spending more than 15 minutes on any problem if you are stuck (take as much time for the actual implementation as you need). If some concept is completely new to you, then try to watch some video or read up on it. These problems may appear directly, but even if they don't they form a nice basis for the pattern recognition required for DSA. Your goal is to kind of internalize these problems at least. It should not take more than 15 minutes to solve any of these problems after you have practiced them enough. This took me one week of full time effort, and I was familiar with most of these concepts from my undergrad. If you are starting out fresh, it will take longer. Important: This is usually the make-or-break round, so take enough time to prepare well.
If you want to test further, you can try the NeetCode 150 list [link], or pick a few areas where you think you might struggle.
ML Coding can be a bit tricky to prepare for, because it's quite broad. In general, you should know basics of modern architectures thoroughly, and also their APIs in one popular framework (e.g. PyTorch or TensorFlow). LLMs are also useful here. Ask one LLM to generate some code for a type of ML algorithm (say convolutions) and then ask another LLM to find bugs and optimize this code (vectorization, space complexity, etc.) This is extremely useful to get you thinking about the internals of these algorithms. For transformers, the Arena blogs have nice exercises [link]. I was quite familiar with this, and revised them with cs229 and cs231n, so it took around 2-3 days of effort. Important: I would advise against trying to implement everything from scratch. The goal is to refresh quickly. If you really have a lot of time, this is not a bad idea.
In some cases you might even receive a take-home assignment. With or without that, a useful skill is to have a good grip of accepted styles and best practices for code. Like type hints, docstrings, etc. Also some internals of python like copy by reference or copy by value, list comprehensions, etc. These are small things that can help you stand out.
Final rounds are typically HR conversations to discuss logistics, compensation, and team placement. It is important to understand that, you must say that the current firm is your top choice! This is an obvious red flag otherwise.
All the best!