Company interview guide
Microsoft Machine Learning Interview Guide (2026)
Project 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.
Microsoft's machine learning interview loop has 8 reported stages, from recruiter or hiring-team screen to hiring manager, senior leader & behavioral; typical length: 2-8 weeks.
Roles covered: Applied Scientist, ML Engineer, Data Scientist, AI Engineer and Research Scientist. The question bank holds 86 questions reported from Microsoft interviews, 8 of them free to solve in your browser.
- Interview process length
- 2-8 weeks
- Stages
- 8
- Reports linked to bank questions
- 98
- Bank report dates
- Jan 2020–Sep 2026
- Questions in bank
- 86 (8 free)
- Last reviewed
- October 2026
How the Microsoft interview loop works
Usually a short recruiter or HR call (some teams skip it and book the technical screen directly), a 30-60 minute technical screen, then a loop of three to six interviews on Teams, most often back to back on one day and sometimes split over two weeks. A hiring manager, skip-level manager or senior leader often holds the last conversation. Round content is set by the team, and outsourced recruiters frequently share no details, so ask early. Some data science and campus roles start with a timed online test, research scientist loops can add a one-hour job talk, and team matching after the loop has added weeks in a few reports.
The Microsoft interview process, stage by stage
01Recruiter or hiring-team screen
10-30 min call with a recruiter or HR; some teams skip it, and a few data science or campus roles add a timed online test
A short check of background, motivation, level and location. Many reports describe outsourced recruiters who schedule rounds two or three days ahead and share little about content, so push for details. A handful of data science and campus postings send a timed online test (SQL, multiple choice or two coding problems) before any interview.
What it tests
- Fit between your background and the posting
- Why Microsoft and why this team
- Logistics: level, location, timeline, work authorization
How to prepare
- Ask which org and team the role sits in and what each later round covers; content varies a lot by team.
- If you applied to several postings, confirm which one this loop is for.
- For a data science posting, ask whether there is an online test and whether it is SQL, coding or a quiz.
A recruiter or HR call before the technical rounds in 12 ML and data reports, 11 since 2024; online tests in 5, mostly data or campus. The hiring page says some teams add a brief screening call, then 2-4 interviews; reports show 3-6.
02Technical screen
30-60 min on Teams with the hiring manager, a senior scientist or a peer; occasionally on a coding platform
Typically a walk through one project, then a fast set of ML or LLM questions and one short task: an easy to medium coding problem, a small ML function, a classification case study or a GenAI design tied to the team's product. Hiring-manager screens sometimes go mathematical, for example proving a loss is convex. Time is tight; one report had two coding tasks in 30 minutes.
What it tests
- Explaining one project clearly, including why you chose each method
- Bias-variance, precision, recall, F1 and confidence intervals on the spot
- Transformer basics and the practical problems of shipping an LLM feature
- Short coding: counting, string or top-k problems, or text similarity with follow-ups on embeddings
- A small case study or a GenAI design drawn from the team's own product
How to prepare
- Keep your project summary to about five minutes so there is time for the technical part.
- Expect follow-ups on any method you name: if you say bag-of-words, be ready for word2vec and GloVe trade-offs.
- For a hiring-manager screen, read about the team's product; the design prompt is often their daily problem.
Free reported questions for this stage
- Evaluate a Message Classifier: Fit, F1 and UncertaintyClassical MLEasySolve
A distinct first technical round in 33 reports (21 since 2024): 30 or 60 minutes, about ten with the hiring manager. Official pages describe an initial technical interview before the final round, though in general software terms.
03Loop: coding
45-60 min, often after a resume chat; some loops have two coding rounds and a few are mostly coding
Standard data-structure and algorithm problems, mostly easy to medium, with occasional hard ones (dynamic programming, linked lists, a prefix-remainder array trick). Some rounds are low-level design, such as a data structure or LRU cache. Interviewers watch how you move from brute force to an optimal answer and whether you test by hand.
What it tests
- Hashing and counting with clear tie-breaking rules
- Sorting-based array problems, intervals and linked-list manipulation in place
- Trees, graphs and level-order traversal
- Dynamic programming on strings and arrays
- Dry-running your code on edge cases before declaring it done
How to prepare
- State the brute force first, then improve it; reports say the path matters as much as the final code.
- Do not skip coding prep for a scientist role; one applied scientist loop was four algorithm rounds plus a manager round.
- Practise in a plain editor; assume no AI help unless told otherwise, though one 2026 intern loop had a round where AI assistance was explicitly allowed.
Free reported questions for this stage
- Reverse a Linked List, Then Reverse a Sublist In PlaceCoding & AlgorithmsMediumSolve
- Valid AnagramCoding & AlgorithmsEasySolve
Algorithm coding in the loop in 38 ML and data reports, 24 since 2024, mostly easy to medium; official pages say some roles write real code in a compiling editor, in the language you know best.
04Loop: ML coding from scratch
20-60 min inside a screen or loop round, plain Python, numpy or PyTorch
Implement a classic algorithm or evaluation routine without a framework: k-means, logistic regression with its loss, linear regression, a 2D convolution, a small CNN pipeline, tensor operations, or precision and recall for a search API that misbehaves. It shows up in screens as often as in the loop.
What it tests
- Lloyd's k-means with clear stopping rules and empty-cluster handling
- Logistic or linear regression loss and gradient written and vectorized by hand
- Convolution shapes, stride and padding in numpy
- Robust metric code when inputs are missing, short, duplicated or malformed
- Wiring data loading, a model and an evaluation loop end to end
How to prepare
- Be able to write k-means and logistic regression from memory in under 15 minutes.
- Explain convergence and complexity as you code; follow-ups often ask why k-means converges.
- Handle bad inputs explicitly; one round was entirely about an API returning None or the wrong count.
From-scratch ML or numpy coding in 14 reports, 10 since 2024 (k-means twice, logistic regression, 2D convolution, an audio CNN, tensor ops), about half of them in the screen.
05Loop: ML, deep learning & LLM depth
45-60 min with an applied scientist, architect or senior engineer; a project deep dive plus concept questions
The interviewer digs into the details of a past project or paper, sometimes asking how you would design an ablation, then switches to theory. Recent reports cover attention math, BERT versus GPT, normalization layers, gradient problems, multimodal transformers, post-training, RLHF, agents and how to pick an LLM. Research scientist loops at Microsoft Research can add a one-hour job talk.
What it tests
- Attention with Q, K and V, the square-root scaling and multiple heads
- Encoder-only versus decoder-only models and images inside a transformer
- Batch norm versus layer norm, vanishing and exploding gradients
- Metrics, L1 versus L2 and ablation design for a multi-part pipeline
- Choosing an LLM and defining tools for an agent
How to prepare
- For every project, know the baseline, the ablations you ran and what you would try next.
- Answer theory with intuition first, then the formula; reports say interviewers value reasoning over recall.
- Prepare a short view on current topics (RL for LLMs, agents, MCP); senior interviewers bring them up.
Free reported questions for this stage
- Transformer Attention and the 1/sqrt(d_k) ScalingLLMs & GenAIMediumSolve
- How Self-Attention Works, From Intuition to Tensor ShapesLLMs & GenAIMediumSolve
ML, deep learning or LLM questions or a project deep dive in the loop in 40 reports, 25 since 2024; official pages only say to expect questions on past work and on how you would handle role-related tasks.
06Loop: ML & GenAI system design
45-60 min open discussion; sometimes folded into the hiring-manager screen
Recent prompts are mostly LLM products: retrieval-augmented assistants over private documents with citations, slide or ad chatbots, and evaluation pipelines. Classic ML designs also recur, especially recommenders, product search and fraud or abuse detection. Some ML-adjacent roles get a conventional software system design instead.
What it tests
- RAG end to end: ingestion, hybrid retrieval, reranking, grounding and citations
- Evaluating every component and the whole feature before launch
- Latency and serving cost of an LLM feature at large scale
- Ranking and retrieval for search or recommendations
- Safety: prompt injection, harmful content and fairness audits
How to prepare
- Have one RAG blueprint you can adapt to documents, email or slides, with a metric per stage.
- Tie the design to the team's product; interviewers often pick a problem from their own work.
- Agree on scope early; several candidates ran out of time in this round.
An ML, GenAI or system design discussion in 20 reports, 16 since 2024 (RAG, recommenders, search, fraud); the official technical tips page covers classic software and distributed system design, not ML design.
07Loop: statistics, probability & experimentation
45-60 min, or questions inside another round; common in data scientist loops, some applied scientist loops, rare for ML engineers
A round many ML candidates under-prepare for. Expect hypothesis testing, p-values, regression assumptions and R squared, expectation and probability puzzles and, for data scientists, A/B test design, causal questions and SQL.
What it tests
- t-tests, p-values and the central limit theorem explained plainly
- Linear regression: deriving the loss, the assumptions and what breaks when each fails
- A/B test design, sample size and power
- Counting and conditional-probability puzzles
- Basic SQL joins and aggregations
How to prepare
- Be ready to say what a negative R squared means and when the t-distribution applies.
- For experiment questions, name the metric, the unit of randomization and the guardrails before the test.
- Data scientist loops reported interference between treatment and control in ranking tests; have an answer.
Free reported questions for this stage
- Information Gain of a Decision-Tree SplitStatistics & InferenceEasySolve
Statistics, probability, experimentation or SQL questions in 18 reports, 11 since 2024; most are data scientist loops, with a few applied scientist and research engineer ones.
08Hiring manager, senior leader & behavioral
30-60 min with the hiring manager, a skip-level manager or a director; behavioral questions also appear inside technical rounds
Microsoft says it assesses respect, integrity, accountability and growth mindset. Expect questions on working across teams, disagreement, asking for help and feedback, and lately how you use AI tools in your own work, plus a discussion of the role. A senior leader often holds the final conversation, and it can turn technical. A few intern and research loops had no behavioral questions at all.
What it tests
- Moving work forward when another team does not prioritize it
- Handling overlapping scope, conflict and disagreement with a manager
- Learning from feedback and from a decision that turned out wrong
- Ownership and motivation for this team
How to prepare
- Prepare stories about influence without authority; cross-team questions recur.
- Show reflection, not only success: growth mindset questions look for what you changed.
- Ask real questions about the team's roadmap; senior interviewers treat this round as two-way.
Behavioral, culture or hiring-manager rounds in 42 reports, 28 since 2024, usually last; the hiring page lists the values assessed and the tips page suggests STAR answers. Two 2025-2026 intern loops had none.
Questions and themes reported most often
- Deep dive into a past project or paper, opening most rounds (25 reports, loop: ml, deep learning & llm depth)
- Easy to medium algorithm problem, brute force then optimal (16 reports, loop: coding)
- Rapid ML basics: bias-variance, overfitting, precision, recall, F1, ROC, L1 vs L2 (14 reports, technical screen)
- Implement an ML algorithm from scratch: k-means, logistic regression, convolution, a CNN (12 reports, loop: ml coding from scratch)
- Attention, transformers and BERT vs GPT (7 reports, loop: ml, deep learning & llm depth)
- Design a RAG or GenAI assistant and evaluate each stage (6 reports, loop: ml & genai system design)
- Cross-team influence, conflict and asking for help (6 reports, hiring manager, senior leader & behavioral)
- Probability and expectation puzzles, such as riders on a bus with limited seats (5 reports, loop: statistics, probability & experimentation)
What Microsoft asks about
- ML & LLM concepts26%
- General coding22%
- ML / GenAI system design16%
- Behavioral & project deep dive14%
- Stats, probability & experimentation12%
- ML coding from scratch10%
Microsoft interview prep checklist
- Can present two projects in five minutes each and survive detailed follow-ups
- Answered bias-variance, F1, ROC-AUC and L1 vs L2 questions in under a minute each
- Wrote k-means and logistic regression from scratch in numpy without notes
- Solved a set of medium hashing, sorting, linked-list and string-DP problems while explaining trade-offs
- Sketched a RAG assistant with citations, permissions and a metric for every stage
- Explained attention, multi-head attention and BERT vs GPT with tensor shapes
- Prepared stories on cross-team influence, disagreement and learning from a mistake
Tips
- Ask the recruiter which org you are interviewing with and what each round covers; loops differ a lot between teams.
- Prepare two projects in depth: data, baseline, ablations, failures and what you would change. Most rounds start there.
- Drill fast, plain-language answers to core ML questions; screens move through five or six topics in a few minutes.
- Write k-means, logistic regression, softmax cross-entropy and a 2D convolution in numpy until they are routine; this shows up in screens too.
- Keep solving medium algorithm problems even for a scientist role, and talk through brute force before optimizing.
- Have a RAG design ready with evaluation for each stage, and adapt it to the team's product.
- For applied scientist and data scientist roles, revise hypothesis testing, regression assumptions and A/B testing.
- Rounds are often back to back with short breaks; plan for fatigue and keep answers concise.
See the rest of the Microsoft loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the Microsoft loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Microsoft interview FAQ
How many interview rounds does Microsoft have for ML roles?
Candidates report 8 stages: Recruiter or hiring-team screen, Technical screen, Loop: coding, Loop: ML coding from scratch, Loop: ML, deep learning & LLM depth, Loop: ML & GenAI system design, Loop: statistics, probability & experimentation and Hiring manager, senior leader & behavioral. Usually a short recruiter or HR call (some teams skip it and book the technical screen directly), a 30-60 minute technical screen, then a loop of three to six interviews on Teams, most often back to back on one day and sometimes split over two weeks. A hiring manager, skip-level manager or senior leader often holds the last conversation. Round content is set by the team, and outsourced recruiters frequently share no details, so ask early. Some data science and campus roles start with a timed online test, research scientist loops can add a one-hour job talk, and team matching after the loop has added weeks in a few reports.
How long does the Microsoft ML interview process take?
Reported timelines run 2-8 weeks.
What does Microsoft ask in ML interviews?
By reported question mix: ML & LLM concepts 26%; General coding 22%; ML / GenAI system design 16%; Behavioral & project deep dive 14%; Stats, probability & experimentation 12%; ML coding from scratch 10%. The most repeated themes: Deep dive into a past project or paper, opening most rounds; Easy to medium algorithm problem, brute force then optimal; Rapid ML basics: bias-variance, overfitting, precision, recall, F1, ROC, L1 vs L2.
How should I prepare for the Microsoft ML interview?
Ask the recruiter which org you are interviewing with and what each round covers; loops differ a lot between teams. Prepare two projects in depth: data, baseline, ablations, failures and what you would change. Most rounds start there. Drill fast, plain-language answers to core ML questions; screens move through five or six topics in a few minutes.
Are there free Microsoft interview questions?
Yes. LeetML has 86 questions reported from Microsoft interviews, and 8 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- GoogleClassic algorithm rounds plus ML design, ML domain depth and Googleyness, judged by a hiring committee.
- AmazonLeadership Principle questions in every round, standard algorithm coding, and ML breadth, depth and design rounds that differ by track.
- MetaFast classic coding, an AI-assisted coding project and product ML design for recommendations and integrity.
- LinkedInProbability-flavoured coding, ML fundamentals and recommender design built around LinkedIn's own products.
- AppleTeam-run loops built on project deep dives, everyday algorithms, short from-scratch ML coding and product-flavored ML design.
- OpenAIPractical coding, hands-on ML debugging and LLM system design, with a strong mission check.