Company interview guide

Meta Machine Learning Interview Guide (2026)

Fast classic coding, an AI-assisted coding project and product ML design for recommendations and integrity.

Meta's machine learning interview loop has 13 reported stages, from recruiter call to team matching; typical length: 1-4 months.

Roles covered: ML Engineer, Research Scientist, Gen AI Engineer and Data Scientist. The question bank holds 155 questions reported from Meta interviews, 15 of them free to solve in your browser.

Try a free question

Interview process length
1-4 months
Stages
13
Reports linked to bank questions
193
Bank report dates
Jan 2020–Sep 2026
Questions in bank
155 (15 free)
Last reviewed
October 2026

How the Meta interview loop works

A recruiter call, sometimes a 70-90 minute online assessment (mostly research roles), a 45-minute coding screen with two problems, then a virtual onsite of four to six interviews of 45-60 minutes, often split over two days. The most common ML engineer onsite is two coding rounds, one ML system design round and one behavioral round; senior loops more often add a second design round, and since late 2025 one coding round is often AI-enabled. A few candidates were given an extra coding round when the signal was mixed, and a committee review follows the loop. Design and behavioral performance largely set your level, and offers one level below target are common. Team matching follows a pass and can add several weeks.

The Meta interview process, stage by stage

01Recruiter call

15-30 min call

A short call on your background, the ML work you have done and the level you are targeting. It is also where org interest is recorded, for example product ranking, infrastructure, GenAI or research.

What it tests

  • Scope and seniority of past ML work
  • Fit with the org or team that is hiring
  • Logistics: level, location, timeline

How to prepare

  • Ask which rounds your loop contains (AI-enabled coding, general system design, research talk) because loops differ by org.
  • Ask for the official prep guide and practice environment links; the AI-enabled round has its own practice pad.

About 66 candidate reports since 2024 across ML and data-scientist tracks open with a short recruiter call on background, level and org interest; some recruiters only email; Meta's careers blog also describes recruiter prep support.

02Online assessment (some candidates)

70-90 min automated coding test, usually 4 problems of rising difficulty

Not every candidate gets one: it shows up mostly in research scientist reports and a few ML engineer and applied scientist reports, sometimes in place of the live screen. Expect two quick tasks, then longer simulation-style problems where careful reading matters more than tricks.

What it tests

  • Reading long specifications accurately
  • Simulating a stream of commands with correct state
  • Finishing all parts inside the time limit

How to prepare

  • Bank time on the first easy parts and leave most of the clock for the last two.
  • Write small helper functions so later parts can reuse earlier logic.

8 reports from late 2024 to 2026, mostly research scientists plus three ML engineers and one applied scientist; 70-90 minutes and four problems is consistent.

03Technical screen

45 min in CoderPad, two problems, code usually not executed

After a few minutes of introductions you get two problems of about 15-20 minutes each, usually well-known easy or medium types, sometimes with a twist, a follow-up or one hard problem. Research scientist screens are the same pure-coding format. Some screens add one or two behavioral questions, and a few senior tech-lead screens were mostly behavioral.

What it tests

  • Fast, bug-free implementation of common patterns: arrays, strings, trees, heaps, sliding windows
  • Sparse or compressed data handling, such as dot products of sparse or run-length vectors
  • Stating time and space complexity before and after coding
  • Testing your own code by hand without running it
  • Communicating and asking clarifying questions

How to prepare

  • Practise two problems in 40 minutes with a timer and no code execution.
  • Always dry-run one example out loud; several candidates found bugs only after the call.
  • State assumptions explicitly; one candidate passed on code but was marked down for not asking clarifying questions.

Free reported questions for this stage

Roughly 98 ML-role reports since 2024 describe a 45-minute screen with two coding problems; research roles get the same format. A handful add behavioral questions; Meta's prep page also describes a 45-minute screen with an engineer.

04Onsite: coding

Usually 2 x 45 min, two problems per round, same style as the screen; one round may be AI-enabled instead

The same format as the screen with slightly harder material: graph and grid search, heaps, permutations and expression problems. Recruiters often say hard problems are not used, yet hard variants still appear, so practise the known hard types too.

What it tests

  • Grid DFS and BFS: connected components, shortest paths, distances to many targets
  • Heap and k-way merge problems: k smallest, kth pair sum, kth in a sorted matrix
  • Backtracking with duplicates and expression search
  • Exploring an unknown environment through an API with backtracking
  • Handling follow-ups that change constraints

How to prepare

  • Work through the most frequently asked problem types for this company until each takes under 15 minutes.
  • Write every practice problem out at least once; reading solutions is not enough under time pressure.
  • If you get stuck on the first problem, agree a plan with the interviewer so the second still gets time.

Free reported questions for this stage

About 79 ML-role reports since 2024 list two onsite coding rounds of two problems each; recent reports increasingly replace one with the AI-enabled round; Meta's careers blog also describes 45-minute algorithm rounds in real code.

05Onsite: AI-enabled coding

About 60 min in CoderPad with a built-in AI assistant, a small multi-file project with unit tests

Introduced around late 2025; it often, but not always, replaces one of the classic coding rounds, and research scientist loops get it too. You get a themed project, such as a maze solver or a word-list utility, and move through stages: understand the code, fix a bug, then implement and optimise new features; one candidate also had to write test cases. Interviewers differ on how much the assistant may write, so expect to explain your plan and complexity yourself.

What it tests

  • Reading an unfamiliar codebase and its tests quickly
  • Debugging a broken implementation
  • Extending working code over several checkpoints
  • Checking AI-produced code before trusting it
  • Explaining approach and complexity

How to prepare

  • Read the tests first; one candidate spent half the round understanding the spec.
  • Decide the approach yourself, then use the assistant for boilerplate and review, not for the core idea.
  • Run the tests after every change and keep the code readable.

12 reports from 2025-09 to 2026-10, half from research scientists; about a third of loop reports in that period mention it, the rest still list two classic coding rounds.

06ML coding (GenAI and research teams)

45 min, write model code by hand, often without running it

Most ML engineer loops contain no ML coding. GenAI, superintelligence-lab and research teams ask you to write attention or a classic algorithm such as k-means from scratch and discuss its cost and modern variants. Research engineer loops have run a round labelled AI coding that mixes a standard coding problem with an ML problem, separate from the newer assistant-enabled round.

What it tests

  • Scaled dot-product and multi-head causal attention from the formula
  • Time and memory complexity of attention
  • Efficient attention variants such as FlashAttention and linear attention
  • Classic algorithms such as k-means, with your own sanity checks

How to prepare

  • Be able to write masked multi-head attention with shapes in under 15 minutes.
  • No tests are provided; write two tiny checks of your own (shapes, a hand-computed case).

Free reported questions for this stage

8 reports from GenAI, research engineer and research scientist loops since 2024; several ML engineers note that their own loops had no ML coding at all.

07Onsite: ML system design

1-2 x 45 min, open discussion; senior loops more often have two, and some add a general system design round

Product ML design is the round that most often decides level. Common prompts are recommendations for nearby places, news feed or comments, short video, ads and notifications, and detecting harmful or weapon-selling ads. Interviewers want specific signals, labels, funnel stages and metrics, not a generic template.

What it tests

  • Turning a product goal into an ML objective and labels
  • Candidate generation, ranking and re-ranking funnels, and whether each stage needs its own training data
  • Concrete features for users, items and context such as location and time
  • Offline metrics, online A/B tests and guardrails
  • Integrity trade-offs: precision versus recall, human review, adversarial users

How to prepare

  • Budget time: clarification, data and features should not eat the model, evaluation and deployment parts.
  • Name the business intent in integrity prompts; for weapon-selling ads, model the selling signal, not only weapon detection.
  • Have ready blueprints for place recommendation, feed ranking, video recommendation and harmful content detection.

Free reported questions for this stage

About 48 ML-role reports since 2024 name an ML design round (25 more say system design); recommendation and harmful-content prompts dominate. Meta's ML loop guide also lists a 45-minute ML system design interview.

08GenAI and research design

1-2 x 45-60 min, team-specific design and ML depth discussion

GenAI and superintelligence-lab loops replace generic design with work the team really does: pretraining plans, RL post-training, image and multimodal generation, or a chat or coding assistant. Expect questions about recent papers and some loops add an ML fundamentals round on optimizers and scaling laws.

What it tests

  • Data collection and curation for pretraining and post-training
  • Parallelism and memory planning for large training runs, including MoE
  • RL post-training: tasks, reward design and reward weaknesses
  • Modern decoder components: GQA, RoPE, SwiGLU, LoRA
  • Serving and evaluating a chat assistant at scale

How to prepare

  • Read the last year of the team's own papers; one candidate lost the level after missing a detail of a paper.
  • Bring numbers: GPU memory, tokens, throughput and the cost of each design choice.

Free reported questions for this stage

10 reports since late 2024 from GenAI, superintelligence-lab and research engineer loops describe two in-domain design rounds; prompts are specific to the hiring team.

09Research talk and deep dive (research roles)

20-60 min presentation plus paper and project deep dives; research engineer loops often open with a 1-hour hiring-manager deep dive

Research scientist and research engineer loops add a talk on your own work and detailed discussions of your papers, sometimes with an open-ended research problem in your area. Research engineer processes often start with a long hiring-manager interview on your current work, and some research scientist loops end with a hiring-manager round.

What it tests

  • Clarity of motivation, method and results in your own research
  • Defending design choices and alternatives
  • Depth in your domain and awareness of recent work

How to prepare

  • Prepare to explain why you made each modelling choice and why not the obvious alternative.
  • Keep a short version and a long version of the talk ready.

15 research scientist and research engineer reports since 2024: about 7 describe a research talk or paper deep dive, 8 a long hiring-manager or project deep dive.

10Data scientist track: SQL and technical skills

45 min screen and one 45-60 min onsite round, live SQL plus a related product question

Data scientist loops are a separate track. The screen pairs one or two SQL queries with a product case on the same scenario, and the onsite repeats a technical SQL round; one 2026 report describes it as AI-native, with the SQL largely written through an assistant.

What it tests

  • Joins, CTEs, window functions and conditional aggregation
  • Rates, retention and histograms over event logs
  • Tying a query result to a product decision

How to prepare

  • Practise writing correct SQL quickly while talking through edge cases such as nulls and duplicate events.
  • Expect the product follow-up to use the same tables as the query.

Free reported questions for this stage

21 data-scientist reports from 2025-2026 describe live SQL in the screen and an onsite technical round; the screen nearly always pairs it with a product case.

11Data scientist track: analytical execution and reasoning

2 x 45-60 min: statistics, probability and experiments; then a product case

Analytical execution covers probability, hypothesis tests and A/B test design, sometimes with a z-score by hand. Analytical reasoning is a product case on metrics, diagnosis of a metric drop and trade-offs for a Meta surface.

What it tests

  • Bayes rule and expected-value probability questions
  • Experiment design, bias and interference in social products
  • Choosing success and guardrail metrics for a feature
  • Structured diagnosis of a metric drop

How to prepare

  • Use one fixed structure for every metric-drop question: definitions, segments, internal changes, external causes.
  • Know the z-test and confidence interval formulas well enough to compute them by hand.

Free reported questions for this stage

15 data-scientist reports since 2025 describe analytical execution (stats, probability, A/B tests) and analytical reasoning (product case) onsite rounds.

12Behavioral

45 min, 6-10 questions with follow-ups; PhD candidates get a PhD-focused version

A fast round that covers many questions, each with detailed follow-up. It weighs heavily in leveling, so the scope of your stories matters as much as the delivery.

What it tests

  • Most impactful project and your own part in it
  • Conflict, pushback and working with a difficult person
  • Ambiguity, changing scope and tight deadlines
  • Feedback given and received, failures and growth
  • Mentoring and leadership at the target level

How to prepare

  • Prepare 6-8 stories and keep each under two minutes; 45 minutes goes quickly.
  • Lead with measurable impact and scope that matches the level you want.
  • Research candidates should have a crisp answer to why a PhD and how it applies to the role.

Free reported questions for this stage

About 90 reports since 2024 across all tracks include a 45-minute behavioral round; research candidates often get a PhD-focused version. Meta's ML loop guide also describes it.

13Team matching

One or more 30-45 min chats with hiring managers after the onsite

After a positive loop you talk with managers of teams with open headcount; one candidate had four such chats, and interns were asked to rank teams. This step is informal but can take several weeks, longer during reorganisations.

What it tests

  • Interest in and fit with the team's problems
  • Questions you ask about the work

How to prepare

  • Prepare a short list of teams and what you want to work on.
  • If nothing happens for a while, ask your recruiter whether you are waiting for a decision or for a match.

7 reports since 2024 (ML engineers, ML interns and data scientists) describe team-matching chats after the onsite; most reports stop at the onsite outcome.

Questions and themes reported most often

  • Proudest project with deep follow-ups, plus conflict and a difficult coworker (10 reports, behavioral)
  • Nearby place recommendations for a user on the move (6 reports, onsite: ml system design)
  • End-to-end recommender for ads, pages or posts (6 reports, onsite: ml system design)
  • Dot product of sparse, run-length or too-large-for-memory vectors (5 reports, technical screen)
  • Detect harmful or weapon-selling ads and posts (5 reports, onsite: ml system design)
  • Largest island or connected component in a grid (5 reports, onsite: coding)
  • Why a PhD, and research experience questions (4 reports, behavioral)
  • News feed, home feed or comment ranking (4 reports, onsite: ml system design)

What Meta asks about

  • General coding40%
  • ML system design24%
  • Behavioral14%
  • GenAI design & ML depth9%
  • ML coding5%
  • SQL, stats & product (DS track)8%

Meta interview prep checklist

  • Solved two medium problems in 40 minutes without running code, with dry runs
  • Can write sparse and run-length dot products, grid island search and heap k-way merges from memory
  • Completed a timed AI-assisted session on an unfamiliar multi-file project
  • Can present a nearby-place recommender and a harmful-ads detector end to end in 45 minutes
  • Have 6-8 behavioral stories with measurable impact and follow-up details
  • For GenAI or research roles: wrote multi-head attention by hand and prepared a research talk
  • For data science roles: timed SQL drills and a metric-drop diagnosis framework

Tips

  • Coding volume matters most: drill the company's most frequently asked problem types until two mediums fit in 40 minutes.
  • Practise without running code, and finish every solution with a spoken dry run and complexity.
  • Rehearse the AI-enabled format on a small multi-file project: read tests, fix a bug, extend, verify; it now appears in many ML engineer and research scientist loops.
  • Prepare four ML design blueprints: place recommendation, feed ranking, short-video recommendation and harmful content detection.
  • In design rounds, give specific features and labels and keep 15 minutes for modelling, evaluation and launch.
  • Treat behavioral and design as the leveling rounds; size your stories to the level you want.
  • For GenAI or research teams, add attention from scratch, LLM training and post-training design, and the team's recent papers.
  • Data scientists should prepare SQL speed, probability by hand and a fixed metric-diagnosis structure instead of algorithms.

See the rest of the Meta loop

The later stages, the questions reported for them and how to prepare are in Pro.

Unlock with Pro

Practice the Meta loop

Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.

Meta interview FAQ

How many interview rounds does Meta have for ML roles?

Candidates report 13 stages: Recruiter call, Online assessment (some candidates), Technical screen, Onsite: coding, Onsite: AI-enabled coding, ML coding (GenAI and research teams), Onsite: ML system design, GenAI and research design, Research talk and deep dive (research roles), Data scientist track: SQL and technical skills, Data scientist track: analytical execution and reasoning, Behavioral and Team matching. A recruiter call, sometimes a 70-90 minute online assessment (mostly research roles), a 45-minute coding screen with two problems, then a virtual onsite of four to six interviews of 45-60 minutes, often split over two days. The most common ML engineer onsite is two coding rounds, one ML system design round and one behavioral round; senior loops more often add a second design round, and since late 2025 one coding round is often AI-enabled. A few candidates were given an extra coding round when the signal was mixed, and a committee review follows the loop. Design and behavioral performance largely set your level, and offers one level below target are common. Team matching follows a pass and can add several weeks.

How long does the Meta ML interview process take?

Reported timelines run 1-4 months (1-4 months, often about 2).

What does Meta ask in ML interviews?

By reported question mix: General coding 40%; ML system design 24%; Behavioral 14%; GenAI design & ML depth 9%; ML coding 5%; SQL, stats & product (DS track) 8%. The most repeated themes: Proudest project with deep follow-ups, plus conflict and a difficult coworker; Nearby place recommendations for a user on the move; End-to-end recommender for ads, pages or posts.

How should I prepare for the Meta ML interview?

Coding volume matters most: drill the company's most frequently asked problem types until two mediums fit in 40 minutes. Practise without running code, and finish every solution with a spoken dry run and complexity. Rehearse the AI-enabled format on a small multi-file project: read tests, fix a bug, extend, verify; it now appears in many ML engineer and research scientist loops.

Are there free Meta interview questions?

Yes. LeetML has 155 questions reported from Meta interviews, and 15 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources