Meta Machine Learning Engineer Interview Questions (2026)

How many rounds are in the Meta Machine Learning Engineer interview loop?

The full loop typically has five to six rounds: a recruiter screen, a hiring manager or technical phone screen with one coding problem, and an onsite loop of two coding rounds, one ML system design round, and one behavioral round. Some roles add a second ML design or ML theory round. The end-to-end process takes four to six weeks from first contact to offer.

What does the Meta ML system design round actually test?

Meta wants to see whether you can architect a production ML pipeline from requirements to deployment. A typical prompt is 'design a News Feed ranking system' or 'design a click-through-rate prediction model for ads.' Interviewers specifically evaluate five areas: problem framing and metric definition, candidate generation and retrieval, feature engineering at scale, model selection and offline evaluation, and online evaluation plus retraining cadence. Describing a model in isolation without addressing data freshness, serving latency, and A/B testing frameworks will not pass.

How hard is the coding round for Meta ML engineers compared to a standard SWE loop?

Meta expects ML engineers to code like strong software engineers — the bar is not relaxed. Both coding rounds use LeetCode-style problems at medium-to-hard difficulty. Common patterns include BFS/DFS on graphs, dynamic programming, two-pointer sliding window, and tree traversals. Python is accepted and common; interviewers focus on correctness, edge-case handling, and clean structure. You may also receive ML-adjacent coding problems such as implementing a mini gradient descent loop or writing vectorized NumPy operations.

What is Meta's behavioral round and which values does it probe?

Meta's behavioral round is a dedicated 45-minute session, not a short add-on at the end of a technical round. It maps to Meta's six published core values: Move Fast, Build Awesome Things, Be Direct and Respect Your Colleagues, Focus on Long-Term Impact, Live in the Future, and Meta/Metamates/Me. Interviewers probe conflict resolution, decisions made under ambiguity, times you shipped something despite incomplete information, and cross-team collaboration. Strong answers are quantified (user impact, latency improvement, business metric) — naming a value as an adjective without a concrete outcome does not score well.

How does Meta level ML engineers, and what does compensation look like at each level?

Meta levels ML engineers on the standard software engineering ladder: E3 (new grad), E4 (mid-level), E5 (senior), E6 (staff), E7 (principal). Based on Levels.fyi crowdsourced data, median total compensation for US-based MLE roles is approximately $331K at E4, $472K at E5, and $790K at E6. The jump from E5 to E6 is driven almost entirely by RSU grants, not base salary. Leveling is determined during the loop, not at negotiation — your target level should be stated explicitly in the recruiter call.

What happens after the Meta onsite — does Meta have a hiring committee?

Meta does not use Google-style hiring committees. Instead, your interviewers submit scorecards and the hiring manager makes a provisional decision, which is then reviewed by a recruitment coordinator and sometimes a bar-raiser. If you pass, you enter a Team Match phase where you speak with specific teams before an offer is finalized. Team Match can take two to six weeks and gives you some influence over which ML domain (ads, feed, integrity, generative AI) you land in.

Does Meta ask ML theory questions in a dedicated round?

Not always, and this is one of the ways Meta differs from Google. Meta rarely runs a standalone ML theory round. Instead, theory questions appear inside the ML system design round or are woven into coding rounds as follow-up questions. Expect bias-variance trade-offs, handling class imbalance in ads data, precision-recall trade-offs, gradient boosting versus deep learning choice rationale, and embedding-based retrieval concepts. The context is always practical — 'why would you choose this model for this use case at Meta's scale.'

What is the AI-enabled coding round and does it apply to ML engineer roles?

Meta began rolling out an AI-enabled coding round in 2025 for some engineering roles. In this format, you are permitted to use an AI coding assistant during the problem, and the evaluation shifts toward your ability to review, verify, and debug AI-generated code rather than write every line from scratch. As of 2026 this round is not universal across all MLE openings, so confirm with your recruiter whether it will be part of your specific loop. Preparation is the same — deep understanding of algorithms and data structures is necessary to spot incorrect AI output.

How should I prepare for the Meta ML engineer interview in four to five weeks?

Weeks one and two: solve 60–80 LeetCode problems focused on graphs, trees, dynamic programming, and hash maps. Practice talking through your reasoning out loud. Week three: review ML system design by building end-to-end designs for a news feed ranker, an ads CTR predictor, and a content recommendation engine — covering candidate retrieval, feature engineering, model choice, offline and online evaluation, and retraining. Week four: prepare six to eight behavioral stories using STAR format mapped to Meta's core values, each with a quantified outcome. Week five: do two to three mock full loops under timed conditions and review Meta's published ML interview prep guide on metacareers.com.

Landing a Machine Learning Engineer role at Meta means clearing a process that is simultaneously more coding-intensive than most ML-specific loops and more focused on real production systems than a pure software engineering interview. Meta’s ML engineers own models end-to-end — from offline training pipelines to real-time serving infrastructure — and the interview is calibrated to verify that you can operate at that scope. This guide covers the 2026 loop structure, what each round is actually measuring, how expectations differ by level, worked examples of the specific question types you will face, and a four-to-five-week prep plan.

The Meta MLE interview loop from recruiter call to offer

The process runs across five to six rounds and typically takes four to six weeks from first contact to a verbal offer. Each stage is independent, but interviewers read each other’s feedback before the final decision.

1. Recruiter screen (20–30 minutes)

The recruiter validates basic fit: ML specialization area (ranking, NLP, computer vision, generative AI, integrity), target level, current and expected compensation, and timeline. This is also where your target level gets anchored — Meta levels ML engineers on the E3–E7 ladder, and the difference between an E4 and E5 offer is roughly $141K in median total annual compensation according to Levels.fyi data. Do not leave level ambiguous; ask the recruiter which level the role is posted at and whether there is flexibility.

2. Technical phone screen (45–60 minutes)

A Meta engineer — often in a related ML domain — conducts this round over video with a shared coding editor. Expect one algorithmic problem at LeetCode medium difficulty. The interviewer is watching you solve in real time and grading both correctness and how clearly you communicate your reasoning. This round is also used to assess whether you clear the coding bar before committing to a full onsite — a weak performance here ends the process.

3. Onsite / virtual onsite loop (four to five rounds, each 45 minutes)

This is the core of the process. Rounds are scheduled on the same day or across two consecutive days for virtual formats. A standard E5 MLE loop looks like this:

• Two algorithm coding rounds
• One ML system design round
• One behavioral round (Jedi round in Meta’s internal vocabulary)
• Occasionally: a second ML design or ML theory discussion

Each interviewer files an independent scorecard before any debrief. Scores are calibrated on a rubric; a strong signal in one round does not compensate for a failing signal in another.

4. Team Match (two to six weeks)

If your scores clear the bar, you enter Team Match rather than immediately receiving an offer. You speak with three to five teams across Meta’s ML orgs — which include Ads, Feed and Notifications, Watch, AI Products (Meta AI assistant), and Integrity. You can express preferences and teams can choose not to match you, but passing the loop scores is the hardest gate; most candidates who reach Team Match find a team. Compensation is set by level and is largely non-negotiable during Team Match, though RSU refresh timing and bonus percentages have some flexibility.

What Meta uniquely evaluates

Meta’s MLE interviews have three characteristics that distinguish them from Google’s or Amazon’s ML loops.

Coding is not relaxed for ML engineers. Unlike some companies that treat ML engineers as primarily data-science roles, Meta explicitly expects production-level software engineering. Both coding rounds use the same problem difficulty and rubric as the software engineering track. Candidates who are strong ML practitioners but rusty on graph traversal or dynamic programming fail the coding rounds at a significant rate.

System design is deeply specific to Meta’s product surface. The ML system design prompt will almost always reference a Meta product context — News Feed ranking, ads click-through prediction, Instagram Reels recommendation, or content integrity classifiers. Interviewers know these systems from the inside. A generic textbook ML pipeline answer scores poorly; you are expected to show awareness of Meta’s specific scale (billions of daily active users, sub-100ms serving latency requirements, multi-billion impression ad auctions per day).

The behavioral round carries real weight. At many companies, behavioral rounds are pass/fail on a low bar. At Meta, the Jedi round is a full independent scorecard that can alone determine whether you are hired at E5 or down-leveled to E4. It maps directly to Meta’s six core values, and the interviewers are looking for specific evidence of impact, directness, and long-term thinking — not just cultural pleasantness.

Coding rounds: question types and approach

Meta’s two coding rounds each contain one problem, occasionally followed by a harder variant or a complexity analysis discussion. The most commonly reported topic clusters are:

Graphs (BFS/DFS): Number of islands, word ladder, clone graph, minimum cost path in a grid. These appear frequently because Meta’s social graph and content graph are central to its products — graph fluency signals domain relevance.

Trees and recursion: Lowest common ancestor, serialize and deserialize a binary tree, path sum variants. Interviewers push on edge cases (null nodes, single-child paths) to test for completeness.

Dynamic programming: Longest increasing subsequence, coin change variants, edit distance. DP problems are used to assess whether you can recognize subproblem structure and articulate the recurrence.

Hash maps and sliding window: Substring with concatenation of all words, longest substring without repeating characters. These appear in rounds focused on string and array manipulation.

Sample coding question and approach

Question: Given a grid of cells where each cell is either land (1) or water (0), return the number of distinct islands, where an island is defined as a connected group of land cells. Two cells are connected if they are horizontally or vertically adjacent.

Approach outline:

1. Iterate through the grid. When you encounter an unvisited land cell, increment the island count and trigger a BFS or DFS to mark all reachable land cells as visited.
2. Use a visited set or mutate the grid in place (discuss the trade-off with the interviewer — mutating avoids O(m×n) extra space but is destructive).
3. Time complexity: O(m×n) where m and n are grid dimensions. Space: O(min(m,n)) for the BFS queue in the worst case.
4. Edge cases: empty grid, grid of all water, single-cell grid.

Narrate each decision out loud. Meta interviewers explicitly grade communication quality alongside correctness.

ML system design round: what you need to walk through

The ML system design round is 45 minutes. You will receive a prompt like one of these — all of which have been reported by candidates:

• Design a News Feed ranking system for Facebook
• Design a click-through rate prediction model for the Meta Ads auction
• Design an Instagram Reels recommendation system
• Design a content integrity classifier that detects policy-violating posts at scale
• Design a people-you-may-know friend recommendation feature

Interviewers evaluate five areas: problem framing, training data, feature engineering, modeling, and evaluation and deployment. The scoring rubric from Meta’s own published ML prep guide on metacareers.com confirms these five areas explicitly.

How to structure your answer

Problem framing (5 minutes): Define the business objective in terms of a measurable ML target. For News Feed ranking, the business objective (maximize meaningful engagement) translates to a ranking problem over a candidate set. Define your optimization target metric — something like expected dwell time or interaction probability weighted by post type — and name any guardrail metrics (do not reduce content diversity below threshold X).

Candidate generation (5 minutes): At Meta’s scale, you cannot run a heavy ranker over all available content. Describe a two-stage pipeline: fast retrieval (approximate nearest neighbor over embeddings, collaborative filtering, or a lightweight logistic regression) to reduce candidates from millions to thousands, then a heavier ranker. Interviewers push on why you chose each retrieval method.

Feature engineering (10 minutes): This is where Meta-specific knowledge matters. For feed ranking, features fall into three groups: user features (engagement history, demographic signals, device type), content features (post type, author affinity, content embeddings), and context features (time of day, session recency). Discuss how you handle feature freshness — some features (user-author affinity) can tolerate hourly staleness; others (trending topic signal) cannot.

Modeling (10 minutes): For ranking and CTR prediction, Meta’s production systems have historically used gradient-boosted trees (GBDT) for tabular feature sets and deep neural networks with embedding layers for sparse categorical features, often combined in a Wide & Deep or DLRM-style architecture. Name these specifically. Discuss trade-offs: GBDT trains faster and is more interpretable; deep models handle cross-feature interactions better at scale. For integrity classifiers, discuss multi-modal inputs and label noise.

Evaluation and deployment (10 minutes): Describe offline metrics (AUC-ROC for ranking, log loss for CTR, normalized discounted cumulative gain for ordered ranking) and their limits — AUC can be high while the model still misranks the top-1 slot. Describe A/B test design: holdout split, minimum detectable effect, guard rails. Discuss retraining cadence and data freshness requirements. Mention shadow deployment and gradual rollout. Interviewers want to see that you treat the model as a living system, not a one-time artifact.

Sample system design answer excerpt

Prompt: Design a CTR prediction model for Meta’s ads auction.

Candidate response (modeling section): “At Meta’s auction scale — tens of billions of ad impressions per day — I’d use a two-tower neural network with sparse embeddings for user and ad IDs combined with dense feature layers for behavioral features like past click rates. The towers produce user and ad embeddings that are dot-producted and passed through a prediction head. For training I’d use binary cross-entropy on click/no-click labels from the prior 30 days of impression data, with temporal stratification to avoid future leakage. Serving latency is the main constraint: I’d precompute and cache user embeddings on a TTL of roughly 15 minutes, and keep ad embeddings refreshed hourly. For the final scoring step in the auction, the model runs on CPUs in-datacenter with p99 latency under 10ms. I’d instrument calibration drift with a daily job comparing predicted CTR versus observed CTR by advertiser segment, with an alert threshold at plus or minus 5% relative error.”

This level of specificity — latency numbers, retraining cadence, calibration monitoring — is what differentiates a hiring response from a passing one.

Behavioral round: Meta’s Jedi interview

Meta calls its behavioral round the Jedi interview. It is a standalone 45-minute session, not a few questions tacked onto a technical round. Your interviewer is typically a senior or staff engineer, not an HR partner.

Questions probe six core values:

• Move Fast: “Tell me about a time you shipped under incomplete information and it turned out to be the right call.”
• Build Awesome Things: “What is the most technically ambitious ML project you have worked on? What made it hard?”
• Be Direct and Respect Your Colleagues: “Describe a time you disagreed with your manager or a peer about a technical direction. What did you do?”
• Focus on Long-Term Impact: “Tell me about a time you pushed back on a short-term solution because you believed it would create technical debt.”
• Live in the Future: “How are you using or thinking about large language models or foundation models in your current work?”
• Meta/Metamates/Me: “Tell me about a time you put the team’s success above your own recognition.”

Every answer should follow STAR format (Situation, Task, Action, Result) and end with a quantified outcome — a latency improvement, a lift in engagement metric, a reduction in false positive rate, a revenue impact. Interviewers explicitly downgrade answers that lack measurable results, even when the story is compelling.

Prepare six to eight distinct stories. Avoid reusing the same project across multiple questions — interviewers probe for range of experience.

Level and compensation context

Meta’s engineering ladder for ML engineers runs E3 through E7, with E8 and E9 being distinguished engineer and fellow levels that do not hire externally. Practically speaking, most external candidates target E4 (mid-level) or E5 (senior).

Based on Levels.fyi crowdsourced data from 2025–2026 US offers:

The compensation jump from E5 to E6 is approximately $318K annually and is driven almost entirely by RSU grants, not base salary. Pushing for accurate leveling — by demonstrating E5-caliber system design scope and cross-functional leadership during the loop — is the highest-leverage compensation action available to candidates. Behavioral round performance alone has been documented to shift offers from E4 to E5.

Meta also reorganized its AI engineering structure in 2025 and 2026 under Meta Superintelligence Labs, which now encompasses FAIR (fundamental AI research), AI Products, Applied Research, and MSL Infrastructure. ML engineer roles exist across all four organizations, with compensation and scope varying by team. Ask during Team Match which organization the role maps to — FAIR roles tend to emphasize research skills, while Ads and Feed ML roles emphasize systems engineering and shipping velocity.

Four-to-five-week prep plan

Weeks one and two: coding foundations

Solve 60–80 LeetCode problems with emphasis on graphs, trees, dynamic programming, and hash maps. Prioritize problems tagged with patterns common at Meta: BFS/DFS, two-pointer, sliding window, and binary tree manipulation. Practice in Python — it is the most common language in Meta MLE loops. After each problem, write out time and space complexity before looking at solutions.

Week three: ML system design

Build three end-to-end designs: a feed ranking system, an ads CTR predictor, and a content recommendation engine. For each, write out the full pipeline — problem framing, candidate generation, feature engineering, model selection, offline evaluation, A/B test design, and monitoring. Time yourself to 45 minutes per design. Review Meta’s published ML onsite prep guide at metacareers.com — it confirms the five evaluation areas and includes worked examples.

Week four: behavioral preparation

Map your work history to six of Meta’s core values. Write STAR-structured stories for each, ending with a quantified outcome. Practice delivering each story in under three minutes without reading from notes. Record yourself at least once — the most common failure mode in behavioral rounds is running over time or burying the result in hedged language.

Week five: mock full loops and calibration

Do two to three mock full loops under timed conditions — 45 minutes per round with no pausing. Review your performance on each round. Identify the weakest area and spend the final days reinforcing it. If system design is the gap, design two new systems you have not practiced. If coding is the gap, solve five more hard graph problems.

A well-tracked job search helps you manage the multiple active loops that characterize a serious FAANG job search. When you are managing recruiter follow-ups, team match conversations, and competing offers simultaneously, keeping a clear record of where each application stands and what actions are pending is the difference between timing offers correctly and leaving negotiation leverage on the table.