Company interview guide
Google Machine Learning Interview Guide (2026)
Classic algorithm rounds plus ML design, ML domain depth and Googleyness, judged by a hiring committee.
Google's machine learning interview loop has 10 reported stages, from recruiter call and hiring assessment to hiring committee and team match; typical length: 3-12 weeks.
Roles covered: ML Engineer, Applied Scientist, Research Scientist and Data Scientist. The question bank holds 93 questions reported from Google interviews, 15 of them free to solve in your browser.
- Interview process length
- 3-12 weeks
- Stages
- 10
- Reports linked to bank questions
- 111
- Bank report dates
- Dec 2019–Sep 2026
- Questions in bank
- 93 (15 free)
- Last reviewed
- October 2026
How the Google interview loop works
A resume screen (rarely an online assessment first), a recruiter call, one or two screens of 45-60 minutes, then four or five interviews of about 45 minutes, often in one day. Most screens are a coding problem, but since 2025 some teams screen with an ML fundamentals, modeling or ML design round instead. For ML engineers the onsite is usually two coding rounds, one ML system design, one ML domain round and one Googleyness and leadership round; some loops swap a coding round for ML coding. Candidates report about three to five weeks for the interviews themselves. A hiring committee reviews the written feedback, an extra round is sometimes added when it is mixed, and team matching can add several weeks.
The Google interview process, stage by stage
01Recruiter call and hiring assessment
15-30 min call; a few candidates first take an online multiple-choice hiring assessment or a short online coding test
A short conversation about your background, the role and the loop. Recruiters usually send a preparation guide and list the round types; a few candidates are asked to complete an online assessment before any interview.
What it tests
- Why Google and why this role
- Whether your experience fits a generalist MLE loop or a specialised track
- Logistics: level, location, timeline, interview language
How to prepare
- Ask which rounds you will get (coding, ML coding, ML design, ML domain, research) and whether any use a specific language such as C++.
- Ask how much time you can have to prepare; candidates who accepted a one-week window regretted it.
- Treat the prep guide as a hint, not a contract; reported rounds have differed from it.
A recruiter or HR call opens the process in 19 reports since 2024; an online assessment shows up in only 2. Google's hiring page also describes recruiter conversations and an optional assessment.
02Technical phone screen
1-2 x 45-60 min; usually one algorithms problem plus follow-up in a shared doc or plain editor; some teams use an ML round instead
Most ML candidates get a standard coding screen at medium or hard level, often with a follow-up that changes the constraints. Since 2025 some teams screen with an ML fundamentals or LLM round, a modeling case, an ML design problem or a separate Googleyness screen instead, so ask your recruiter which one you have.
What it tests
- Picking the right data structure quickly (heaps, hash maps, prefix sums, intervals)
- Reaching an optimal-complexity solution, not just brute force
- Writing working code without autocomplete or a compiler
- Explaining time and space complexity and handling the follow-up
How to prepare
- Practise in a plain document with no syntax highlighting; you will not be able to run the code.
- State the brute force in one sentence, then move to the better approach; some interviewers time-box the optimal answer.
- Walk through two or three test cases by hand before saying you are done.
19 ML-role reports since 2024 describe a screen: about 12 were a single coding problem, 5 (all 2025-2026) were an ML fundamentals, modeling or ML design screen; Google's own hiring guides name a phone screen stage.
03Onsite: coding rounds
Usually 2 (sometimes 3) x 45 min of algorithms and data structures, in your choice of language
The core of the Google loop and where many ML candidates fail. Graph and tree traversal dominate the reports, followed by data structure design with a twist and harder dynamic programming or probability problems. Interviewers look for clean, bug-free code and self-written tests, often add a scale follow-up, and sometimes add an extra coding round when feedback is mixed.
What it tests
- BFS and DFS on implicit graphs, grids and link graphs
- Designing a data structure with a modified rule (for example an LRU cache with pinned keys)
- Dynamic programming and expected-value reasoning under time pressure
- Interval scheduling with priorities or weights
- Edge cases, input validation and unit tests
How to prepare
- Make graphs and trees your first priority; they are the most frequently reported topic.
- Finish with your own test cases, including empty input, cycles and duplicates; reports name missed edge cases as the reason for failing.
- Prepare a scale follow-up for every graph problem: caching expensive distance calls, pruning, or batching.
- Expect Python fundamentals questions if you code in Python, such as recursion depth limits and hashing of mutable objects.
DSA coding rounds appear in 25 ML-role reports since 2024, usually two onsite. Google's hiring and interview-prep pages also describe coding exercises and technical interviews.
04Onsite: ML coding
45-60 min in Python with numpy or PyTorch; replaces a coding round in some loops
Reported in a growing share of 2024-2026 loops, across engineer, research and applied roles. Candidates implement a core model component from scratch, most often multi-head attention, a full transformer block or the language-model head, and answer questions about state-of-the-art variants as they go. A few reports instead describe building a small model live, such as a linear regression or a CNN classifier.
What it tests
- Multi-head attention with correct reshapes, scaling and masking
- Transformer block details: pre-norm, residuals, gated feed-forward such as SwiGLU
- Turning hidden states into vocabulary logits and sampling from them
- Small NLP utilities such as a tokenizer, or a simple model trained live
- Explaining shape and cost of every operation you write
How to prepare
- Be able to write attention, a pre-norm block and an LM head in numpy and in PyTorch without looking anything up.
- Say the tensor shape out loud after every line; it catches most bugs before the interviewer does.
- Know the common variants (grouped-query attention, rotary positions, SwiGLU) well enough to explain why they are used.
Free reported questions for this stage
- Multi-Head Causal Self-Attention From ScratchML CodingHardSolve
- Pre-Norm GPT Transformer Block Forward PassML CodingHardSolve
- Multi-Head AttentionML CodingMediumSolve
- Implement Grouped-Query AttentionML CodingHardSolve
- Byte-Pair Encoding Tokenizer from ScratchML CodingHardSolve
9 reports since 2024 describe an ML coding round: 5 ML engineer loops plus research scientist, research engineer and applied scientist loops; attention or transformer code is the most common task.
05ML domain and fundamentals
45-60 min discussion; can be a phone screen or an onsite round, often starting from your resume
An interviewer picks topics from your background and goes deep: model choice, losses, regularization, metrics and, increasingly, transformers and the LLM training pipeline. It can be a screen or an onsite round. Some versions are a modeling case on a small dataset, or an open design question, with quick-fire theory questions mixed in.
What it tests
- Explaining the transformer and the full LLM lifecycle with trade-offs
- Classical fundamentals: logistic regression, Naive Bayes, bagging versus boosting, bias and variance
- Regularization methods and when each helps
- Choosing metrics for imbalanced or multi-class problems
- Framing a modeling case: EDA, features, data split, generalization
How to prepare
- Prepare a short, structured answer for each core concept, then expect the follow-up to go one level deeper than you planned.
- Know every model and loss on your resume well enough to derive it; the round often starts there.
- In modeling cases, talk about how you would split the data to avoid leakage before you talk about models.
Free reported questions for this stage
- Macro, Micro and Weighted F1 for a Ticket RouterClassical MLMediumSolve
- Multiplying Replicated Matrices on Two GPUsML System DesignMediumSolve
- Deriving Loss Gradients by Hand: Softmax Cross-EntropyDeep LearningEasySolve
16 reports since 2024 describe an ML domain, applied ML or fundamentals round, often starting from the resume; 3 of the 2025-2026 ones ran it as a screen; Google's interviewing guide lists role-related knowledge.
06Onsite: ML system design
About 45-60 min, open discussion; often described as the biggest filter
Most reports describe recommendation systems (a real-time feed, watch-next videos, a bandit choosing one of a few candidates) and trust and safety systems (copyright matching, spam, fraud, near-duplicate videos) at very large scale; a few were GenAI or question-answering systems. Interviewers expect a production design: serving, latency, monitoring and iteration, not just a model.
What it tests
- Candidate generation, ranking and re-ranking for a large catalog
- Cold start, exploration and online learning from immediate feedback
- Latency versus accuracy trade-offs, caching and near-line serving
- Detection at scale with embeddings, approximate search and human review
- Writing down the loss for each model and how it is optimized
How to prepare
- Have a full recommender blueprint ready, including cold start and monitoring; shallow answers on those were the stated reason for at least one rejection.
- Some interviewers ask for exact loss functions and optimizer update rules, so practise writing them for ranking and retrieval models.
- Prepare a separate blueprint for matching or abuse detection: fingerprints or embeddings, approximate nearest neighbour search, thresholds and an appeal flow.
Free reported questions for this stage
- Contextual Bandit Choosing One of Four CandidatesML System DesignHardSolve
An ML system design round appears in 17 ML-role reports since 2024, usually one onsite round; recommenders lead, then abuse or matching systems. One team used it as a screen.
07Research tracks: talk, research interviews, LLM depth
Research scientist: a 45-60 min job talk plus 2-3 research interviews; research-engineer style LLM roles: deep technical conversations with the hiring manager
Research scientist loops add a talk on your work and research interviews that mix questions about your experimental choices with general ML and probability. LLM-focused research engineer roles can skip coding at first and probe training math in depth. Formats vary by team, so ask which apply.
What it tests
- Defending design decisions in your own research and proposing extensions
- KL divergence, importance sampling, perplexity and the VAE objective, including why LLM training uses them
- Post-training methods and how to detect reward hacking
- Large-scale training: parallelism, stability and evaluation
- Probability and math questions alongside coding
How to prepare
- Build the talk around two or three decisions you can defend, and prepare an answer on how the work connects to the team.
- For LLM roles, be able to derive and explain each quantity, not just define it; reports say memorised answers were not enough.
- Research coding rounds are still algorithm problems; do not skip them.
Free reported questions for this stage
- VAE Objective: ELBO Gap, Gaussian KL and Posterior CollapseDeep LearningMediumSolve
12 research-track reports since 2024 describe a job talk, research interviews or an LLM-depth conversation with the manager; older research-scientist reports show the same talk-plus-interviews loop.
08Data scientist track: statistics, analysis and product
Two gating interviews (statistical knowledge, data analysis and intuition), then further rounds on applied ML, coding, experiments and behavior
Data scientist and product analyst loops lean on statistics and clear communication. In several reports the first two interviews decide whether you continue; later rounds cover experiments, product metrics, Python and SQL, and some ML. Many practice problems here come from older question lists, so treat them as coverage rather than exact repeats.
What it tests
- Hypothesis tests, p-values and confidence intervals explained plainly
- Regression assumptions, multicollinearity and estimator properties
- A/B test design: metric choice, sample size, novelty and long-term effects
- Product metrics and diagnosing a drop or low adoption
- Probability puzzles and SQL
How to prepare
- Practise explaining each statistical idea to a non-technical stakeholder; communication is scored explicitly.
- For every experiment question, name the guardrail and long-term metrics before computing anything.
- Know which interview you are in; candidates report studying for one format and getting another.
Free reported questions for this stage
- Is a Coin Biased After 550 Heads in 1,000 Flips?Statistics & InferenceEasySolve
- Confidence Intervals for Logistic Regression ParametersStatistics & InferenceMediumSolve
- Breaking a Stick into a TriangleProbabilityEasySolve
12 data scientist reports since 2024 describe statistics, analysis, experiment or product rounds; 3 say two stats and analysis interviews gate the rest. Two said the prep guide did not match.
09Googleyness and leadership
One round of about 45 min, plus a project deep dive inside other rounds; some teams run it as a screen
A structured behavioral round on collaboration, ambiguity, failure and leadership without authority. Interviewers in every round also score these qualities, and a resume or dissertation deep dive often appears in the ML rounds.
What it tests
- Conflict with a teammate, manager or supervisor and how it ended
- A failure and what you changed afterwards
- Handling changing priorities and many requests at once
- Going beyond the brief and pushing back on the status quo
- Depth behind your proudest project
How to prepare
- Prepare four or five STAR stories with numbers; loosely structured answers were cited in rejections.
- Expect a disagreement question even in research loops, for example with a PhD supervisor.
- Prepare how you would respond when a manager doubts a model that won its A/B test; it was asked in a behavioral round.
A behavioral or Googleyness round appears in 31 reports since 2024 across all tracks; 2 ML loops ran it as a separate screen. Google's own interviewing guide describes behavioral and hypothetical leadership questions.
10Hiring committee and team match
No interview with you; then calls with hiring managers until one team agrees
Written feedback goes to a hiring committee of people who did not interview you, and it can ask for an extra round. Candidates who pass the general bar then talk with hiring managers; this can happen before or after the committee and can take weeks. Level is set here and down-levelling happens.
What it tests
- Consistency of signal across all rounds
- Fit between your experience and a specific team's work
How to prepare
- Use team-match calls to show depth in the team's area; treat them as short interviews.
- Keep other processes moving; candidates report waiting in team match after passing every round.
6 reports since 2024 mention the hiring committee or team matching, one with an extra round after committee review; a research report was down-levelled. Google's own pages say hiring committees decide; team match is not described.
Questions and themes reported most often
- Research talk followed by questions on experimental choices (16 reports, research tracks: talk, research interviews, llm depth)
- Graph or grid traversal with BFS or DFS, then a scale follow-up (13 reports, onsite: coding rounds)
- Conflict, failure and changing priorities, told in STAR form (9 reports, googleyness and leadership)
- Deep dive on your proudest project or dissertation (9 reports, googleyness and leadership)
- Explain the transformer architecture and its modern variants (8 reports, ml domain and fundamentals)
- Design a recommendation system: real-time feed, watch-next or a bandit slot (7 reports, onsite: ml system design)
- Statistics round: p-values, hypothesis tests and regression assumptions (6 reports, data scientist track: statistics, analysis and product)
- Quick-fire ML fundamentals: regularization, boosting, metrics, overfitting (5 reports, ml domain and fundamentals)
What Google asks about
- General coding (DSA)34%
- ML system design18%
- ML & LLM concepts17%
- Behavioral & deep dive12%
- ML coding10%
- Stats & probability9%
Google interview prep checklist
- Solved ten graph and tree problems in a plain editor, each with self-written tests
- Can implement an LRU cache variant and a weighted interval scheduler without hints
- Wrote multi-head attention, a pre-norm transformer block and an LM head from scratch
- Can sketch an end-to-end recommender with cold start, serving latency and monitoring
- Can explain the transformer and the LLM training pipeline with trade-offs
- Prepared STAR stories for conflict, failure and changing priorities
- Know the round list and coding language for your specific loop
Tips
- Do not under-prepare algorithms; for most ML engineer loops at least half the rounds are standard coding, and they decide many outcomes.
- Drill graphs and trees first, then intervals, heaps and dynamic programming, always in a plain editor with no autocomplete.
- End every coding answer with your own test cases; interviewers mark down untested code even when the idea is right.
- Have a production-grade recommender design ready, covering cold start, latency, monitoring and the exact losses you would train.
- Be able to write multi-head attention and a transformer block from scratch; ML coding rounds are becoming more common.
- Know your resume projects to the level of losses, metrics and failed alternatives; ML domain rounds start there.
- Ask your recruiter for the round list and any language requirement; tracks and teams vary more than the standard loop suggests.
- Prepare four or five STAR stories, and plan for team matching to take weeks after you pass.
See the rest of the Google loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the Google loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Google interview FAQ
How many interview rounds does Google have for ML roles?
Candidates report 10 stages: Recruiter call and hiring assessment, Technical phone screen, Onsite: coding rounds, Onsite: ML coding, ML domain and fundamentals, Onsite: ML system design, Research tracks: talk, research interviews, LLM depth, Data scientist track: statistics, analysis and product, Googleyness and leadership and Hiring committee and team match. A resume screen (rarely an online assessment first), a recruiter call, one or two screens of 45-60 minutes, then four or five interviews of about 45 minutes, often in one day. Most screens are a coding problem, but since 2025 some teams screen with an ML fundamentals, modeling or ML design round instead. For ML engineers the onsite is usually two coding rounds, one ML system design, one ML domain round and one Googleyness and leadership round; some loops swap a coding round for ML coding. Candidates report about three to five weeks for the interviews themselves. A hiring committee reviews the written feedback, an extra round is sometimes added when it is mixed, and team matching can add several weeks.
How long does the Google ML interview process take?
Reported timelines run 3-12 weeks.
What does Google ask in ML interviews?
By reported question mix: General coding (DSA) 34%; ML system design 18%; ML & LLM concepts 17%; Behavioral & deep dive 12%; ML coding 10%; Stats & probability 9%. The most repeated themes: Research talk followed by questions on experimental choices; Graph or grid traversal with BFS or DFS, then a scale follow-up; Conflict, failure and changing priorities, told in STAR form.
How should I prepare for the Google ML interview?
Do not under-prepare algorithms; for most ML engineer loops at least half the rounds are standard coding, and they decide many outcomes. Drill graphs and trees first, then intervals, heaps and dynamic programming, always in a plain editor with no autocomplete. End every coding answer with your own test cases; interviewers mark down untested code even when the idea is right.
Are there free Google interview questions?
Yes. LeetML has 93 questions reported from Google interviews, and 15 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- Google DeepMindAlgorithmic coding and broad maths, stats and ML fundamentals, then team, research and culture conversations.
- MetaFast classic coding, an AI-assisted coding project and product ML design for recommendations and integrity.
- WaymoGraph-heavy coding, numpy ML coding and driving-flavored ML design, plus a research talk for scientists.
- OpenAIPractical coding, hands-on ML debugging and LLM system design, with a strong mission check.
- 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.
- AmazonLeadership Principle questions in every round, standard algorithm coding, and ML breadth, depth and design rounds that differ by track.
Official sources
- How we hire (Google Careers, official)
- Interview tips (Google Careers, official)
- Google: A guide to structured interviewing for better hiring practices (official)
- Google: Effective interviewer training for better candidate experiences (official)
- Google: How to create a positive experience for candidates (official)