Candidates describe Meta's AI Research Scientist interviews as a mix of LeetCode-style coding, in-depth research discussion or presentation, and some ML, design and behavioral rounds. Formats vary by team.
$190K
Avg. Base Comp
$390K
Avg. Total Comp
2-6 rounds
Typical Rounds
Not reported
Process Length
Candidates describe the Meta AI Research Scientist interview as coding plus research depth, and several say the coding mattered more than they expected for a research role. Reports mention easy-to-medium LeetCode-style problems, with some harder ones, such as topological sorting, graph or DFS problems, tree traversal and merging sorted arrays. One candidate was asked to implement k-means from scratch in CoderPad and didn't finish. Another didn't advance past a phone screen built around a graph bipartition problem. Practice explaining complexity and edge cases aloud, because candidates report follow-ups on both.
The research side varies by team. Some candidates presented their work for about 40-45 minutes and were interrupted with questions on loss functions, failed experiments and alternatives. Others report questions on LoRA, RLHF, multimodal LLM design, or why they wouldn't use DiT. One candidate was asked to propose a full research plan with a model, baselines and benchmarks. Another says the discussion centered on how well their work matched the team's direction. Expect to defend your own design decisions rather than summarize papers.
The rest of the loop is less consistent. Some candidates report ML system design, such as binary classification, and distribution-mismatch questions. Others report OS and networking fundamentals, or design rounds tied to the interviewers' research areas. Others found ML content minimal. Behavioral rounds were described as fairly standard, with questions on PhD challenges and proudest projects. Outcomes in these reports include offers, rejections and a declined offer, and several candidates mention slow scheduling or rescheduled final days.
Synthesized from 15 candidate reports by our editorial team.
Had an interview recently?
Share your experience. Unlock the full guide.
Real interview reports from people who went through the Meta process.
I went through two rounds in total, and the split between them was the most useful thing I learned going in. The first was a coding interview. It covered standard LeetCode-style problems, but it also had an ML component where I worked with PyTorch, so I'd make sure you can write model and training code without leaning on a reference, not just solve algorithm puzzles. The second round was a research discussion about my projects, my background, and my areas of expertise. It was more of a conversation than a quiz, and it's where I spent most of my energy.
I had heard before the interview that a strong research background could help a candidate even if the coding went imperfectly. I can't say how much that weighed in my case, but the research round clearly centered on how well my work matched the research direction of the team. The more concretely I could connect my past work to what they were trying to build, the more natural the discussion felt.
Overall the process felt relaxed. I received an offer and accepted it.
Prep tip from this candidate
Prepare to write PyTorch code under interview conditions alongside standard coding problems, since the ML portion was part of the coding round. Spend time before the research discussion mapping each of your projects to the team's research direction, because that alignment appeared to drive the outcome.
Share your own interview experience to unlock all reports, or subscribe for full access.
Sourced from candidate reports and verified by our team.
Topics based on recent interview experiences.
Featured question at Meta
Select the 2nd highest salary in the engineering department
| Question | |
|---|---|
| Merge Sorted Lists | |
| Experiment Validity | |
| Decreasing Comments | |
| Scrambled Tickets | |
| 500 Cards | |
| Friendship Timeline | |
| Weighted Keys | |
| Button AB Test | |
| P-value to a Layman | |
| Swipe Precision | |
| Nearest Common Ancestor | |
| Groups of Anagrams | |
| Find the Missing Number | |
| Using R Squared | |
| Recurring Character | |
| Impression Reach | |
| Bank Fraud Model | |
| Lazy Raters | |
| Radix Addition | |
| Find the First Non-Repeating Character in a String | |
| Network Experiment Design | |
| Twenty Variants | |
| Reservoir Sampling Stream | |
| Fill None Values | |
| Booking Regression | |
| Production Model Monitoring | |
| Hurdles In Data Projects | |
| Reducing Error Margin | |
| Find Bigrams |
Synthesized from candidate reports. Individual experiences may vary.
Candidates report an initial recruiter call or reach-out covering background, research experience, interest in the role, preferred teams and timeline. One candidate also describes a Zoom intro about the research talent the team was seeking. Treat it as a chance to clarify team fit, since some later rounds appear team dependent.
Several candidates report a first technical screen with two LeetCode-style problems, typically easy to medium, though one 45-minute screen was described as medium to hard. Others report a CodeSignal assessment with four questions. One candidate was given a graph bipartition problem in CoderPad. Explain edge cases and complexity aloud, since candidates say follow-ups focus on reasoning.
Candidates report research deep dives or presentations lasting about 40-45 minutes, with interruptions probing design choices, failed experiments and alternatives. Topics reported include LLM training, RLHF, LoRA and multimodal model design. One candidate was asked for a complete research plan with baselines and benchmarks. Questions may also be tied to the interviewers' research areas.
Later rounds differ across reports. Some candidates describe ML system design (binary classification, train/test distribution mismatch), ML fundamentals reasoning, or design rounds matched to an interviewer's field. Others mention PyTorch coding, sparse-matrix problems, IoU implementation, or OS and networking questions. A few found ML content light, so prepare broadly rather than assuming one format.
Candidates describe the behavioral round as fairly standard, covering past projects, collaboration, PhD difficulties and handling ambiguity. Full-loop days are reported as several back-to-back interviews, and some candidates mention slow feedback, rescheduling or missing interviewers. Outcomes in the reports range from offers to rejections, so keep your calendar flexible.
