Company interview guide
LinkedIn Machine Learning Interview Guide (2026)
Probability-flavoured coding, ML fundamentals and recommender design built around LinkedIn's own products.
LinkedIn's machine learning interview loop has 8 reported stages, from recruiter call to data scientist track: sql, product and experiments; typical length: 4-8 weeks.
Roles covered: AI Engineer, ML Engineer, Applied Scientist and Data Scientist. The question bank holds 42 questions reported from LinkedIn interviews, 3 of them free to solve in your browser.
Owned by Microsoft. Microsoft acquired LinkedIn in December 2016. LinkedIn runs its own hiring and interview loop.
- Interview process length
- 4-8 weeks
- Stages
- 8
- Reports linked to bank questions
- 48
- Bank report dates
- Aug 2017–Oct 2026
- Questions in bank
- 42 (3 free)
- Last reviewed
- October 2026
How the LinkedIn interview loop works
A recruiter call, sometimes a call with the hiring manager, then a one-hour technical screen. Up to 2023 the screen was usually half coding and half ML concept questions; since 2025 most reported screens fit a few behavioral questions, one coding or ML coding problem and a short ML design into the hour. The virtual onsite is usually five 60-minute rounds. Two coding rounds appear in nearly every report; the rest is one to three ML or system design rounds, a behavioral or hiring-manager round and, in some loops, an ML foundations round. Interns and data scientists often start with an online assessment. Feedback can take weeks, and several candidates waited a month or more.
The LinkedIn interview process, stage by stage
01Recruiter call
15-45 min call; the company says a video call with the hiring manager or a team member may follow
A screen of your background, why LinkedIn and which team fits. Recruiters usually explain the later rounds, and one 2025 candidate also had a separate hiring-manager round before the onsite in which the manager described the round types.
What it tests
- Why LinkedIn, and why this team
- How your past ML work maps to the role and level
- Logistics: location, timeline, level
How to prepare
- Ask how the screen hour is split (coding, ML questions, design, behavioral) so you can pace it.
- Ask which product area the team owns; design rounds are often drawn from it.
Recruiter or HR call first in 34 candidate reports across ML and data science roles (shared stage); the company's hiring page describes the recruiter call and an optional manager or team call.
02Online assessment (interns and data scientists)
Timed online test, about 75-80 min: two coding problems for ML interns, SQL plus some Python or multiple choice for data scientists
Intern and data scientist candidates often take an automated test before any live interview. Intern tests are two coding problems where the naive solution times out, or a short probability-in-code task; data science tests are mostly SQL joins on given tables. No experienced ML engineer report since 2023 mentions one.
What it tests
- Optimizing a brute-force solution so it passes time limits
- Turning a uniform random generator into a uniform integer range
- SQL joins and aggregation across several tables
How to prepare
- Read the given schema carefully; one candidate found table names in the prompt that did not match the data.
- Submit a working solution first, then optimize; one candidate passed with a slow first answer.
Online test before interviews in 10 candidate reports from 2021-2023: six ML intern or early-career ML reports and four data scientist reports; none from experienced ML hires since 2024.
03Technical phone screen
60 min, shared editor, often two interviewers; recent screens: brief behavioral, one coding or ML coding problem, short ML design; older screens: coding plus an ML concept quiz
The coding part very often has a probability twist: turn one random source into another with exact probabilities, or sample from a weighted distribution. Several 2026 screens used ML coding instead, such as debugging a numpy logistic regression or writing pseudocode for an embedding-based training pipeline. Most recent screens end with a 15-20 minute ML design, so time management matters as much as correctness.
What it tests
- Rejection sampling and bias removal with a proof that the output is exactly uniform
- Weighted sampling with prefix sums and binary search, plus unnormalized weights
- Deriving and simplifying the logistic regression gradient while fixing buggy code
- Short, correct answers to ML concept questions (overfitting, regularization, metrics)
- Fitting behavioral, coding and design into one hour
How to prepare
- Know the expected number of calls for every rejection scheme you write; interviewers ask.
- Carry the logistic regression gradient all the way to X^T (h - y) / m; one candidate was marked down for stopping early.
- Keep each behavioral answer to about two minutes so the coding and design parts get their time.
- Run a dry run of your code before saying you are done; some interviewers expect it to execute.
One-hour technical screen in 37 ML-role reports (15 since 2024). Nine had a sampling-in-code problem; 7 of the 8 reports since 2025 added a mini ML design; older ones paired coding with an ML quiz.
04Onsite: coding
2 x 60 min, data structures and algorithms in any language; some loops label one of them AI coding
Medium to hard algorithm problems, each extended with follow-ups. Recent examples include total covered length of intervals, first by sorting and then with a segment tree under adds and removes; a word-ladder search that gains a neighbour cache and a changing dictionary; a weighted-die sampler with unnormalized and skewed weights that ended in a uniform-point-in-a-disk question; and a monotonic stack problem. A 2024 candidate said the AI coding round felt like a regular algorithm round.
What it tests
- Interval merging, then a segment tree or similar structure that supports updates
- Graph search (BFS) with caching and thread-safe access to shared data
- Heaps, sliding windows, two pointers and monotonic stacks
- Handling follow-ups on scale, skewed inputs and concurrency
How to prepare
- Practise writing a segment tree from memory; knowing the idea without having coded it cost one candidate.
- When a follow-up mentions updates, ask whether you need true concurrency or just a lock around the shared read.
- Leave time to test edge cases; candidates report running out of time after long explanations.
Free reported questions for this stage
- Uniform Point in the Unit Disk Without RejectionProbabilityMediumSolve
- K-Way Merge with a Min-HeapCoding & AlgorithmsMediumSolve
Two onsite coding rounds in 17 ML-role reports (6 since 2024), algorithmic, sometimes with a probability follow-up; the company's AI-use page mentions coding assessments but gives no count or format.
05ML system design
1-3 x 60 min onsite; most recent screens also include a 15-20 min mini version
Most prompts are LinkedIn products: the feed, job recommendations and click models, Learning course recommendations from job-search activity, skill inference, companies or pages to follow and LLM-drafted recruiter messages. Interviewers probe data and labels, feature engineering and ranking methods in detail. One 2025 AI-track loop asked infrastructure designs instead: a metrics monitoring platform and a distributed key-value store.
What it tests
- Defining labels from impressions, handling position bias and avoiding leakage
- Extracting features from job posts, profiles and skills, including with LLMs
- Pointwise, pairwise and listwise learning to rank and when each fits
- Cold start for new members and new postings
- Combining ML forecasts with rule-based controls (for example ad budget pacing)
How to prepare
- Prepare a recommender blueprint with LinkedIn entities: members, jobs, skills, courses, posts.
- Mention how you would extract structured features from job post text early; one interviewer had to prompt for it.
- In the mini version inside a screen, stay high level and cover features before models.
- If you are on an AI-track team, still review one classic distributed systems design.
Free reported questions for this stage
- Designing a Multi-Tenant Distributed Key-Value StoreML System DesignHardSolve
ML or system design in 20 ML-role reports (13 since 2024), onsite or as a mini design in the screen; recommendation problems on LinkedIn surfaces dominate, one loop used infrastructure design.
06Onsite: ML foundations & probability
60 min discussion with some math, in some onsites only; sometimes starts from a small modelling scenario
A breadth round that moves quickly across practical ML and, at LinkedIn more than most companies, probability. A 2026 report covered calibration, feature selection with neural networks, tokenizers and Adam; older reports add logistic regression derivations, theory proofs and probability puzzles. Two of the five onsite reports since 2024 had this round; the others spent those slots on more design rounds.
What it tests
- When and how to calibrate predicted probabilities, and how to measure calibration
- Feature selection methods and optimizer internals such as Adam's moment estimates
- Logistic regression: the model, its loss and its link to maximum likelihood
- Bayes' rule, expected waiting times and uniform sampling on a disk or sphere
How to prepare
- Prepare crisp one-paragraph answers for calibration, regularization, bagging vs boosting, AUC, batch norm and Adam.
- Write logistic regression and its gradient from scratch in numpy at least once before the loop.
- Practise Bayes' rule with unreliable signals and expected-waiting-time puzzles; one 2022 onsite asked three in one hour.
An onsite ML theory or probability round in 10 ML-role reports, but only 2 since 2024; most support is 2017-2022 and recent loops often replace it with design.
07Behavioral & hiring manager
45-60 min with the hiring manager or a senior engineer; short behavioral blocks also appear in screens; director loops had three managerial rounds
Standard behavioral questions about deadlines, disagreements and mentoring, plus a walk through your most impactful project. At staff level one candidate was given a recommendation project and asked to plan it: milestones, headcount, timeline and cross-team coordination.
What it tests
- Shipping under a tight deadline and what you chose to cut
- Disagreeing with a manager and committing to the outcome
- Mentoring a junior engineer or scientist
- For senior roles: planning and staffing a multi-team ML project
How to prepare
- Have three distinct stories with measurable results ready to tell in two minutes each.
- For staff interviews, practise turning a product idea into phases, roles and a rough headcount.
Behavioral or hiring-manager talk in 25 ML and DS reports (shared stage), 10 since 2024; the company's hiring page describes a manager or team call on your experience, and says to expect follow-ups on your reasoning.
08Data scientist track: SQL, product and experiments
Data scientist loops: SQL and Python screen, case or problem-solving screen, then an onsite of data manipulation, product case, applied statistics or A/B testing, data storytelling and hiring manager
Data science loops replace most algorithm work with SQL on event logs, metric design for LinkedIn features, experiment analysis and presenting insights from a small dataset. Some recruiters ask a basic SQL question on the first call. Nearly all of these reports are from 2017-2021, so confirm the current shape with your recruiter.
What it tests
- SQL with joins, self-joins, window functions and aggregation over activity logs, sometimes repeated in pandas
- Defining success metrics for a LinkedIn feature
- Attributing a metric change to a launch when no clean experiment exists
- Basic statistics and A/B test design
- Telling a clear story from a small sample of data
How to prepare
- Practise SQL on page-view style logs with timestamps; ambiguous schemas are reported.
- Use difference-in-differences style reasoning when one platform changed and another did not.
SQL, product case, statistics or A/B rounds in 19 data scientist reports, consistent with each other but nearly all from 2017-2021; no detailed data science loop report since 2024.
Questions and themes reported most often
- ML concept quiz: overfitting, regularization, boosting, AUC, optimizers (15 reports, technical phone screen)
- Recommend jobs, feed posts, companies or pages for LinkedIn members (10 reports, ml system design)
- SQL on member activity logs plus a product metric case (8 reports, data scientist track: sql, product and experiments)
- Exact uniform 0 to 6 (or rand7) from a biased or fair coin (6 reports, technical phone screen)
- Logistic regression from scratch: code it, debug it, derive its gradient (6 reports, technical phone screen)
- Probability puzzles: Bayes with unreliable reports, waiting times, geometric sampling (5 reports, onsite: ml foundations & probability)
- Sample an index from a weighted die, then handle unnormalized or skewed weights (4 reports, technical phone screen)
- Deadline, disagreement, mentoring and impactful-project stories (4 reports, behavioral & hiring manager)
What LinkedIn asks about
- Coding & algorithms27%
- ML / recsys system design25%
- Probability & sampling18%
- ML fundamentals13%
- ML coding8%
- Behavioral & leadership9%
LinkedIn interview prep checklist
- Wrote an exact uniform 0-6 generator from a coin of unknown bias and stated its expected number of calls
- Implemented weighted sampling with binary search and discussed unnormalized and skewed weights
- Coded and debugged logistic regression in numpy with the gradient fully simplified
- Can design a LinkedIn job or feed recommender, including labels, leakage and learning to rank
- Can explain calibration, feature selection, tokenizers and Adam in a few sentences each
- Solved an interval problem with a segment tree that supports adds and removes
- Prepared deadline, disagreement and mentoring stories with concrete results
Tips
- Drill random-source conversions until you can write them and prove exactness quickly: biased to fair, fair to 0-6, rand5 to rand7, uniform float to an integer range.
- Know weighted sampling cold: prefix sums, binary search, the alias method for skewed or huge distributions, and the cost of rejection.
- Write logistic regression, its loss and its simplified gradient by hand; it shows up as coding, debugging and theory.
- Build one recommender design around LinkedIn entities and rehearse labels, features, ranking loss, cold start and leakage.
- Expect quick follow-ups on every coding problem (caching, updates, concurrency, scale) and keep time for them.
- Practise one-hour mocks that combine behavioral questions, a coding problem and a mini design, since recent screens use that shape.
- Refresh practical ML breadth (calibration, feature selection, tokenizers, Adam, regularization, metrics); it is asked in screens even when the onsite has no theory round.
- Ask your recruiter or hiring manager which round types you will get; some candidates were told the themes in advance.
See the rest of the LinkedIn loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the LinkedIn loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
LinkedIn interview FAQ
How many interview rounds does LinkedIn have for ML roles?
Candidates report 8 stages: Recruiter call, Online assessment (interns and data scientists), Technical phone screen, Onsite: coding, ML system design, Onsite: ML foundations & probability, Behavioral & hiring manager and Data scientist track: SQL, product and experiments. A recruiter call, sometimes a call with the hiring manager, then a one-hour technical screen. Up to 2023 the screen was usually half coding and half ML concept questions; since 2025 most reported screens fit a few behavioral questions, one coding or ML coding problem and a short ML design into the hour. The virtual onsite is usually five 60-minute rounds. Two coding rounds appear in nearly every report; the rest is one to three ML or system design rounds, a behavioral or hiring-manager round and, in some loops, an ML foundations round. Interns and data scientists often start with an online assessment. Feedback can take weeks, and several candidates waited a month or more.
How long does the LinkedIn ML interview process take?
Reported timelines run 4-8 weeks.
What does LinkedIn ask in ML interviews?
By reported question mix: Coding & algorithms 27%; ML / recsys system design 25%; Probability & sampling 18%; ML fundamentals 13%; ML coding 8%; Behavioral & leadership 9%. The most repeated themes: ML concept quiz: overfitting, regularization, boosting, AUC, optimizers; Recommend jobs, feed posts, companies or pages for LinkedIn members; SQL on member activity logs plus a product metric case.
How should I prepare for the LinkedIn ML interview?
Drill random-source conversions until you can write them and prove exactness quickly: biased to fair, fair to 0-6, rand5 to rand7, uniform float to an integer range. Know weighted sampling cold: prefix sums, binary search, the alias method for skewed or huge distributions, and the cost of rejection. Write logistic regression, its loss and its simplified gradient by hand; it shows up as coding, debugging and theory.
Are there free LinkedIn interview questions?
Yes. LeetML has 42 questions reported from LinkedIn interviews, and 3 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- MetaFast classic coding, an AI-assisted coding project and product ML design for recommendations and integrity.
- GoogleClassic algorithm rounds plus ML design, ML domain depth and Googleyness, judged by a hiring committee.
- MicrosoftProject deep dives in most rounds, quick ML and LLM fundamentals, standard coding, from-scratch ML code and GenAI design, in a loop shaped by the hiring team.
- PinterestRapid-fire ML fundamentals, edge-case-heavy coding and embedding-centric recommendation design.
- IndeedA take-home modeling task, then rounds on coding, probability and statistics, and ML that stay close to classic data science.
- SnapClassic coding plus fast, broad ML fundamentals and recommendation-heavy ML design, with values questions in every round.