Meta-Style Problems

Reading time: ~45 min | Interview relevance: Critical (Meta/Instagram/WhatsApp) | Roles: Machine Learning Engineer, Software Engineer (ML), Research Scientist, Data Scientist, ML Infrastructure Engineer at Meta

Meta interviews move fast. Where Google wants you to discuss three approaches before coding, Meta wants you to pick the right approach immediately and execute cleanly. Where Google gives you 45 minutes to explore a problem, Meta often packs two problems into 45 minutes.

Meta's ML interviews are deeply product-oriented. Every ML system design question ties back to a specific Meta product: News Feed ranking, Instagram Reels recommendation, Marketplace fraud detection, WhatsApp spam filtering. If you cannot connect your ML knowledge to product impact, you will struggle in Meta interviews.

This list of 30 problems is calibrated to Meta's actual interview style based on publicly reported experiences from 2023-2025. The emphasis is on speed, product ML, recommendation systems, and practical execution.

Meta Interview Structure (ML Roles)

Round	Duration	Evaluation Focus	Meta-Specific Notes
Coding 1	45 min	DSA (2 problems)	Two problems, not one. Speed is essential
Coding 2	45 min	ML coding or system coding	Implementation of ML algorithms or data systems
ML System Design	45 min	End-to-end ML system	Always tied to a Meta product
Behavioral	45 min	Impact, collaboration, growth	Meta values: Move Fast, Be Bold, Focus on Long-Term Impact

:::tip What Makes Meta Interviews Unique

1. Speed. Meta coding rounds typically include two problems in 45 minutes. You have ~20 minutes per problem.
2. Product ML. Every system design ties to a real Meta product (Feed, Reels, Marketplace, Ads).
3. Recommendation systems. Meta is fundamentally a recommendation company. Expect recommendation-focused questions.
4. Execution over exploration. Meta values candidates who can identify the right approach quickly and execute.
5. Impact stories. Behavioral rounds focus heavily on measurable business impact. :::

Section 1: Coding Problems (12 Problems)

Meta coding rounds are fast-paced. You will likely face two problems in 45 minutes. The first is usually medium, the second is medium-hard. Speed is critical.

Fast-Execution DSA

#	Problem	Difficulty	Time	Key Pattern	Meta Lens	Why Meta Asks This
1	LRU Cache	Medium	15 min	Hash map + DLL	Execute fast, no bugs	Feature caching in recommendation serving
2	Binary Tree Right Side View	Medium	12 min	BFS level-order	Quick traversal thinking	Tree structures in content taxonomies
3	Add and Search Word (Trie + DFS)	Medium	18 min	Trie with wildcard	Data structure selection	Search autocomplete, content matching
4	Random Pick with Weight	Medium	15 min	Prefix sum + binary search	Probability thinking	Weighted sampling for ad selection, A/B testing
5	Minimum Window Substring	Hard	20 min	Sliding window + hash map	Optimal solution directly	Text analysis, content matching
6	Alien Dictionary	Hard	22 min	Topological sort from constraints	Graph modeling	Ranking order inference, dependency detection

ML-Adjacent Coding

#	Problem	Difficulty	Time	Key Pattern	Meta Lens	Why Meta Asks This
7	Implement Batch Gradient Descent	Medium	15 min	Vectorized computation	NumPy fluency	Core training loop for any model
8	Implement AUC-ROC from Sorted Predictions	Medium	15 min	Threshold sweep	Metric implementation	Primary metric for ranking and classification models
9	Implement Feature Hashing (Hashing Trick)	Medium	15 min	Hash function + modular indexing	Sparse feature encoding	High-cardinality features in ad/content ranking
10	Implement Online Learning Update (SGD with Decay)	Medium	15 min	Gradient update + learning rate	Real-time model updates	Online learning for fresh ranking signals
11	Implement Stratified Sampling	Medium	12 min	Proportional selection	Data handling	Balanced training data for imbalanced classes
12	Implement Cosine Similarity for Sparse Vectors	Easy	10 min	Sparse dot product	Efficient computation	Content similarity, embedding comparison

:::warning Meta Coding Speed Benchmarks

• Must solve a medium problem in 12-18 minutes (not 25-30)
• Must solve a medium + hard pair in under 40 minutes
• Must write bug-free code on the first pass (limited time for debugging)
• Must handle edge cases without being prompted (empty input, single element, overflow)

If you consistently take >20 minutes for medium problems, focus on speed drills before your Meta interview. :::

Section 2: ML System Design (12 Problems)

Meta ML system design is always product-oriented. The interviewer will name a specific Meta product and ask you to design the ML system behind it. You must demonstrate product intuition alongside ML expertise.

Core Product ML

#	Problem	Difficulty	Time	Product	Key Challenge	What Meta Evaluates
13	Design News Feed Ranking	Hard	45 min	Facebook Feed	Multi-objective optimization (engagement, time well spent, integrity)	Understanding of ranking at massive scale with competing objectives
14	Design Instagram Reels Recommendation	Hard	45 min	Instagram Reels	Cold-start for new content, short-video engagement modeling	Exploration-exploitation for creator ecosystem health
15	Design People You May Know (PYMK)	Medium	35 min	Facebook Social	Graph features + ML ranking, privacy constraints	Combining social graph structure with ML scoring
16	Design Facebook Marketplace Item Ranking	Medium	35 min	Marketplace	Location-aware ranking, seller quality, fraud detection	Multi-signal ranking with trust and safety
17	Design Instagram Explore Feed	Medium	40 min	Instagram Explore	Interest discovery, topic diversity, freshness	Balancing personalization with content discovery
18	Design WhatsApp Spam Detection	Medium	35 min	WhatsApp	Encrypted messages (no content access), behavioral signals	ML without content features; metadata and behavioral modeling

Ads & Monetization ML

#	Problem	Difficulty	Time	Product	Key Challenge	What Meta Evaluates
19	Design Meta Ads Targeting System	Hard	45 min	Meta Ads	User interest modeling, privacy (no cross-app tracking post-ATT)	Ads relevance without invasive tracking; privacy-preserving ML
20	Design Ad Auction and Bid Optimization	Hard	45 min	Meta Ads	Auction mechanics, pacing, budget optimization	Revenue maximization with advertiser satisfaction constraints

Integrity & Safety ML

#	Problem	Difficulty	Time	Product	Key Challenge	What Meta Evaluates
21	Design a Hate Speech Detection System	Medium	40 min	Content Integrity	Multilingual, context-dependent, adversarial robustness	Precision vs. recall tradeoff for content moderation at scale
22	Design a Fake Account Detection System	Medium	35 min	Account Integrity	Behavioral signals, graph analysis, evasion detection	Network-based features, temporal patterns, adversarial actors

ML Infrastructure

#	Problem	Difficulty	Time	Product	Key Challenge	What Meta Evaluates
23	Design a Real-Time Feature Pipeline	Hard	45 min	ML Platform	Online feature computation, consistency with offline	Feature freshness vs. consistency; real-time aggregations
24	Design a Model Serving System for 1M QPS	Hard	45 min	ML Platform	Ultra-low latency, high throughput, model versioning	Batching, caching, model optimization (quantization, distillation)

:::note Meta ML Design Framework Meta interviewers expect this structure (move quickly through each):

1. Problem Clarification (2 min): What product? What metric? What scale?
2. Metrics (3 min): Business metric (revenue, engagement) + ML metric (AUC, NDCG) + Guardrail metric (integrity, time well spent)
3. System Architecture (5 min): High-level pipeline (data -> features -> model -> serving)
4. Feature Engineering (8 min): User features, item features, context features, interaction features
5. Model Architecture (7 min): Embedding layers, interaction modeling, multi-task heads
6. Training & Evaluation (5 min): Training data, labels, offline eval, online A/B test
7. Serving (5 min): Latency, throughput, model refresh cadence
8. Deep Dive (5 min): Interviewer picks one area; go deep

Meta interviewers value depth in feature engineering and model architecture. Spend extra time there. :::

Section 3: Product Sense & Applied ML (3 Problems)

Meta places high value on product intuition. These problems test whether you can connect ML decisions to product outcomes.

#	Problem	Difficulty	Time	Topic	What Meta Evaluates
25	Engagement metrics are up but user satisfaction is down. Diagnose and fix.	Medium	25 min	Metric Design	Understanding of goodhart's law; proxy metric failure; time well spent vs. engagement
26	Design an experiment to test a new ranking algorithm with potential integrity risks	Medium	25 min	Experimentation	Holdout groups, guardrail metrics, rollback criteria, staged rollout
27	A model improves AUC by 0.5% offline but shows no online improvement. Why?	Medium	20 min	Debugging	Online/offline gap analysis: distribution shift, feature freshness, position bias, label delay

Section 4: Behavioral (3 Discussion Topics)

Meta behavioral rounds focus on impact, speed, and bold decision-making aligned with Meta's values.

#	Topic	Time	Meta Values Tested
28	Tell me about a time you moved fast and broke something. What did you learn?	15 min	Move Fast: willingness to take risks, learn from failures, iterate quickly
29	Describe a project where you had to make a decision with incomplete data	15 min	Be Bold: comfort with ambiguity, bias toward action, iterative approach
30	Tell me about your most impactful ML project. Quantify the impact.	15 min	Focus on Impact: metrics-driven thinking, connecting ML to business outcomes

:::tip Meta Behavioral Preparation Meta values differ from Google's Googleyness:

• Move Fast: Show you bias toward action over analysis paralysis
• Be Bold: Show you take calculated risks rather than always playing it safe
• Focus on Long-Term Impact: Show you think beyond the current quarter
• Build Social Value: Show you consider societal impact of your work
• Be Open: Show you share information and collaborate transparently

Every story should include quantified impact: "This improved CTR by X%, which translated to $Y revenue impact." :::

Meta-Specific Preparation Strategies

1. The Two-Problem Speed Drill

Meta coding rounds have two problems. Practice solving two medium problems back-to-back:

Session	Problem 1 (15 min)	Problem 2 (20 min)	Rest (10 min)
Day 1	Two Sum	Group Anagrams	Review approach speed
Day 2	Valid Parentheses	LRU Cache	Review implementation speed
Day 3	Binary Search	Course Schedule	Review debugging speed
Day 4	#1 (LRU Cache)	#5 (Min Window Substring)	Time yourself
Day 5	#2 (Right Side View)	#6 (Alien Dictionary)	Target: <35 min total

2. Product ML Mapping

For every ML concept, connect it to a Meta product:

ML Concept	Meta Product Application
Collaborative filtering	News Feed ranking, Reels recommendation
Embedding models	User/content embeddings for retrieval
Multi-task learning	Predicting multiple engagement signals (like, share, comment, hide)
Calibration	Ad auction bid optimization
Online learning	Real-time ranking signal updates
Anomaly detection	Fake account detection, platform abuse
Multi-objective optimization	Engagement vs. integrity vs. time well spent
Graph neural networks	Social graph features for PYMK

3. Feature Engineering Depth

Meta interviewers spend significant time on feature engineering. Prepare detailed feature lists for recommendation systems:

User Features:
- Demographics (age bucket, country, language)
- Historical engagement (likes, comments, shares, hides \text{---} last 1h, 1d, 7d, 30d)
- Social graph (friend count, group memberships, page follows)
- Session features (time of day, device, connection type)
- Interest embeddings (learned from historical engagement)

Item Features:
- Content type (text, image, video, link)
- Creator features (follower count, post frequency, integrity score)
- Content embeddings (visual, textual, multimodal)
- Engagement statistics (historical CTR, share rate, hide rate)
- Freshness (time since posted)

Cross Features:
- User-creator affinity (historical interaction between this user and creator)
- Social proof (did friends engage with this content?)
- Topic match (user interest vs. content topic similarity)
- Contextual match (content type preference by time of day)

4-Week Meta Prep Plan

Week	Focus	Problems	Daily Load
Week 1	Coding speed drills	#1-12	2-3 problems/day, timed
Week 2	Product ML design	#13-18	1 design/day
Week 3	Ads/integrity/infra design	#19-24	1 design/day
Week 4	Applied ML + behavioral + mocks	#25-30 + mocks	Mixed daily

Week 1: Coding Speed Drills

Day 1: #1, #2 (LRU Cache, Right Side View \text{---} target 30 min total)
Day 2: #3, #4 (Trie with wildcard, weighted random \text{---} target 35 min total)
Day 3: #5, #6 (Min window substring, alien dictionary \text{---} target 40 min total)
Day 4: #7, #8 (Batch GD, AUC-ROC \text{---} target 30 min total)
Day 5: #9, #10 (Feature hashing, online learning \text{---} target 30 min total)
Day 6: #11, #12 (Stratified sampling, cosine similarity \text{---} target 22 min total)
Day 7: Speed re-run: solve ALL 12 problems, targeting <4 hours total

Week 2: Product ML Design

Day 1: #13 (News Feed ranking \text{---} the canonical Meta problem)
Day 2: #14 (Instagram Reels \text{---} cold-start + engagement)
Day 3: #15 (PYMK \text{---} graph + ML combination)
Day 4: #16 (Marketplace ranking \text{---} location + trust)
Day 5: #17 (Instagram Explore \text{---} discovery + diversity)
Day 6: #18 (WhatsApp spam \text{---} behavioral signals without content)
Day 7: Re-do #13 and #14 from scratch under 45-min time pressure

Week 3: Ads, Integrity, Infrastructure

Day 1: #19 (Ads targeting \text{---} post-ATT privacy challenges)
Day 2: #20 (Ad auction \text{---} mechanism design + ML)
Day 3: #21 (Hate speech detection \text{---} precision/recall for content mod)
Day 4: #22 (Fake account detection \text{---} graph + behavioral)
Day 5: #23 (Real-time feature pipeline \text{---} freshness + consistency)
Day 6: #24 (Model serving at 1M QPS \text{---} optimization + batching)
Day 7: Review all designs; practice feature engineering depth

Week 4: Applied ML + Behavioral + Mocks

Day 1: #25 (Engagement vs. satisfaction diagnosis)
Day 2: #26 (Experiment design with integrity risks)
Day 3: #27 (Online/offline metric gap debugging)
Day 4: #28, #29, #30 (Behavioral stories \text{---} practice out loud)
Day 5: Full mock: 2 coding + 1 ML design + 1 behavioral
Day 6: Full mock: 2 coding + 1 ML design + 1 behavioral
Day 7: Final review; speed-run coding problems one more time

Problem Deep Dives

Problem 13: Design News Feed Ranking

Why Meta asks this: News Feed is Meta's core product. It touches every ML concept: recommendation, ranking, multi-objective optimization, real-time inference, and responsible AI. Your answer reveals your breadth and product intuition.

Architecture:

Content Pool -> Candidate Selection -> Scoring -> Ranking -> Integrity Filter -> User

1. Candidate Selection (~500 candidates from ~1500 eligible)
   - Friends' posts (social graph filtering)
   - Followed pages and groups
   - Suggested content (explore-like discovery)
   - Ads (separate auction winner injection)

2. Multi-Objective Scoring
   - Predict P(like), P(comment), P(share), P(hide), P(report)
   - Each prediction from a separate head on a shared backbone
   - Combine: Score = w1*P(like) + w2*P(comment) + w3*P(share) - w4*P(hide) - w5*P(report)
   - Weights tuned via online experiments

3. Model Architecture
   - User tower: user features + history embeddings
   - Content tower: content features + creator embeddings
   - Interaction layer: cross-attention or factorization machines
   - Multi-task heads: shared bottom, task-specific tops

4. Features (the differentiator in Meta interviews)
   - User: engagement history, time patterns, device, social graph
   - Content: type, creator, age, engagement rate, embedding
   - Context: time, session depth, previous items shown
   - Social: friend engagement signals, social proof

5. Responsible Ranking
   - Integrity classifier demotes harmful content
   - "Time well spent" objective alongside engagement
   - Diversity injection to avoid filter bubbles
   - Creator distribution fairness

6. Serving
   - Score at impression time (real-time)
   - Feature freshness: user features updated every few minutes
   - Model refresh: daily or weekly depending on model size
   - Latency budget: <100ms for full ranking

Problem 24: Design Model Serving at 1M QPS

Why Meta asks this: Meta serves ML predictions for every feed load, ad impression, and content decision. At Meta scale, model serving is an infrastructure challenge as much as an ML challenge.

Architecture:

1. Request Flow
   Client -> Load Balancer -> Feature Assembly -> Model Inference -> Response

2. Feature Assembly (critical path)
   - Online features: fetch from feature store (Redis/Memcache, <5ms)
   - Real-time features: compute on-the-fly (session signals)
   - Batch features: pre-computed, cached
   - Feature assembly budget: <10ms

3. Model Inference
   - Model format: TorchScript, TensorRT, or ONNX
   - Optimization:
     - Quantization (FP32 -> INT8, 2-4x speedup)
     - Distillation (large teacher -> small student)
     - Pruning (remove unimportant weights)
     - Operator fusion (combine layers for fewer kernel launches)
   - Batching: accumulate requests for GPU efficiency
   - Inference budget: <10ms

4. Scaling Strategies
   - Horizontal: more inference servers behind load balancer
   - Model sharding: split model across GPUs for large models
   - Caching: cache predictions for popular items (short TTL)
   - Tiered models: lightweight model for most requests, heavy for high-value

5. Reliability
   - Canary deployment: 1\% traffic to new model, monitor
   - Fallback: if inference fails, serve from lightweight backup model
   - Circuit breaker: if error rate > threshold, route to fallback
   - Graceful degradation: drop optional features under load

Meta vs. Google Comparison

Dimension	Meta Style	Google Style
Coding rounds	2 problems/round (speed)	1 problem + follow-ups (depth)
Approach discussion	Pick and execute fast	Discuss 2-3 approaches first
System design	Product-specific (Feed, Ads)	Generic + scale (Search, YouTube)
Evaluation emphasis	Execution speed, product sense	Clean code, multiple approaches
Behavioral	Impact quantification, boldness	Googleyness, humility, collaboration
Feature engineering	Deep detail expected	Mentioned but less deep
Follow-ups	New problem	Harder constraint on same problem

:::danger Common Reasons Meta Rejects Candidates

1. Too slow on coding. Cannot complete two problems in 45 minutes.
2. Generic system design. Describes a generic recommendation system without product specifics.
3. Shallow feature engineering. Cannot name 20+ features for a recommendation system.
4. No quantified impact. Cannot tie technical decisions to business outcomes.
5. Over-thinking coding. Spends 10 minutes discussing approach when Meta wants you to code.
6. Missing integrity considerations. Does not mention content safety, fake accounts, or adversarial users. :::

Difficulty Distribution

Difficulty	Problems	Count
Easy	#12	1
Medium	#1, #2, #3, #4, #7, #8, #9, #10, #11, #15, #16, #17, #18, #21, #22, #25, #26, #27	18
Hard	#5, #6, #13, #14, #19, #20, #23, #24	8
Behavioral	#28, #29, #30	3

Meta Interview Level Expectations

Level	Coding Speed	ML Design Depth	Product Intuition	Impact Stories
E3 (IC3)	2 mediums in 40 min	Basic ML system	Awareness of product	"I contributed to..."
E4 (IC4)	1 medium + 1 hard in 42 min	Full product ML system	Connect ML to product metrics	"My work improved X by Y%"
E5 (IC5)	2 problems incl. hard in 40 min	Product ML with depth + tradeoffs	Propose product improvements	"I led a project that..."
E6 (IC6)	Clean, optimal, teach-level	Novel system with multi-team impact	Define product direction	"I changed how the team..."

Problem Deep Dives (Continued)

Problem 1: LRU Cache (Meta Speed Version)

Why Meta asks this as a speed problem: At Meta, LRU Cache is not a 25-minute problem. It is a 15-minute warm-up problem. If you spend more than 15 minutes on it, you will not have time for the second problem. This means you must have the implementation memorized and be able to write it without thinking.

Speed Implementation Checklist:

[ ] Declare Node class (key, val, prev, next) \text{---} 30 seconds
[ ] Declare LRUCache with capacity, cache dict, head/tail dummies \text{---} 30 seconds
[ ] Wire head.next = tail, tail.prev = head \text{---} 10 seconds
[ ] Write _remove(node) helper \text{---} 30 seconds
[ ] Write _add_to_front(node) helper \text{---} 30 seconds
[ ] Write get(key) \text{---} 30 seconds
[ ] Write put(key, value) \text{---} 60 seconds
[ ] Test with example \text{---} 60 seconds
Total: ~5 minutes if practiced

Speed Practice Drill:

1. Set a 5-minute timer
2. Implement LRU Cache from scratch (no reference)
3. Test with: put(1,1), put(2,2), get(1), put(3,3), get(2)
4. If it takes >8 minutes, practice again tomorrow

Problem 4: Random Pick with Weight

Why Meta asks this: Weighted random selection is fundamental to ad serving, A/B test traffic allocation, and content sampling. This problem tests whether you understand prefix sums and binary search for probability distributions.

Solution:

import random
import bisect

class WeightedRandom:
    def __init__(self, weights):
        self.prefix = []
        running = 0
        for w in weights:
            running += w
            self.prefix.append(running)
        self.total = running

    def pick(self):
        target = random.random() * self.total
        return bisect.bisect_left(self.prefix, target)

Key Insight: Build a prefix sum of weights. Generate a random number in [0, total_weight). Binary search for where it falls in the prefix array. This gives O(log n) picks after O(n) preprocessing.

Follow-Up Questions:

• What if weights change frequently? (Use a Fenwick tree / BIT for O(log n) updates)
• What if you need to sample without replacement? (Reservoir sampling with weighted version)
• How does this relate to Meta's ad auction? (Weighted sampling for ad candidate selection)

Problem 9: Implement Feature Hashing (Hashing Trick)

Why Meta asks this: Feature hashing is used extensively in Meta's ads and ranking systems to handle high-cardinality categorical features without maintaining a full vocabulary.

Solution:

import hashlib
import numpy as np

def feature_hash(features, n_buckets=2**20):
    """
    Hash feature names + values into a fixed-size vector.
    Handles collision by signed hashing (reduces bias).
    """
    vector = np.zeros(n_buckets)

    for name, value in features.items():
        # Hash to bucket index
        key = f"{name}={value}"
        hash_val = int(hashlib.md5(key.encode()).hexdigest(), 16)
        bucket = hash_val \% n_buckets

        # Signed hashing: use another hash to determine sign
        sign_hash = int(hashlib.sha1(key.encode()).hexdigest(), 16)
        sign = 1 if sign_hash \% 2 == 0 else -1

        vector[bucket] += sign

    return vector

Why Signed Hashing Matters:

• Without sign: hash collisions always add, creating a positive bias
• With sign: collisions cancel on average (E[error] = 0), maintaining unbiasedness
• This is the Weinberger et al. (2009) hashing trick

Interview Discussion Points:

• Tradeoff: more buckets = fewer collisions but more memory
• Feature hashing vs. vocabulary-based encoding: no OOV problem, but collisions lose information
• Used in Vowpal Wabbit, scikit-learn's HashingVectorizer, and Meta's ads stack

Problem 14: Design Instagram Reels Recommendation

Why Meta asks this: Reels is Meta's fastest-growing product and a direct competitor to TikTok. The recommendation system must handle cold-start content, optimize for watch time, and maintain creator ecosystem health.

Architecture:

Content Pool -> Candidate Generation -> Ranking -> Reranking -> Serving

1. Candidate Generation (multiple sources)
   a. Interest-based: user embedding -> ANN search in content embedding space
   b. Social: content engaged by friends/following
   c. Trending: globally popular content in user's locale
   d. Creator-based: new content from followed creators
   e. Exploration: random high-quality content for discovery
   Total: ~500-1000 candidates

2. Ranking Model
   - Architecture: Two-tower for retrieval, cross-attention for ranking
   - Multi-objective prediction:
     P(finish_watch), P(like), P(share), P(comment), P(follow_creator)
     P(not_interested), P(report)
   - Labels:
     - Positive: watch >80\% of video, like, share
     - Negative: skip within 2s, "not interested", report
   - Key feature: watch-through rate, not just click-through rate

3. Cold-Start Handling (Critical for Reels)
   - Content cold-start: embed using visual + audio + text features
   - Exploration allocation: reserve 10-20\% of inventory for new content
   - Creator cold-start: boost new creators' content in explore
   - Progressive rollout: show to small audience, measure engagement, expand

4. Reranking Layer
   - Diversity: no two consecutive reels from same creator
   - Pacing: mix content types (funny, educational, music, news)
   - Creator fairness: minimum impression share for active creators
   - Integrity: filter harmful, misleading, or IP-violating content

5. Engagement vs. Well-Being
   - "Take a Break" reminders after extended sessions
   - Reduce recommendation of divisive content
   - Balance addictiveness with responsible design

Key Metrics:

Metric Type	Metric	Why It Matters
Primary	Daily active users on Reels	User adoption
Engagement	Average watch time per session	Content quality
Content	Watch-through rate (finish watching a reel)	Better than CTR for video
Creator	Creator retention rate	Healthy creator ecosystem
Integrity	% content flagged by users	Platform safety
Guardrail	Session length distribution (not just mean)	Prevent unhealthy usage patterns

Meta Interview Timeline

Understanding Meta's interview process helps with scheduling and preparation:

Stage	Timeline	What Happens
Recruiter screen	Day 0	30-min call; background, motivation, role fit
Technical phone screen	Day 7-14	45-min coding (1-2 problems)
Phone screen result	Day 14-21	Pass/fail notification
Onsite scheduling	Day 21-35	4-5 rounds scheduled over 1-2 days (virtual or in-person)
Onsite interviews	Day 35-50	Full loop: 2 coding + 1 ML design + 1 behavioral
Hiring committee	Day 50-65	Committee reviews packet; may request additional signal
Offer	Day 65-80	Compensation discussion and offer letter

:::tip Meta Timeline Optimization

• Ask the recruiter for the exact interview format (it varies by team)
• Request at least 3 weeks between phone screen and onsite
• If you pass the phone screen, the hardest part is behind you (phone screen has the highest bar)
• Meta often lets you choose your preferred coding language :::

Meta-Specific ML Concepts to Study

Concepts that appear disproportionately in Meta interviews:

Concept	Why Meta Cares	Where It Appears
Multi-task learning	One model predicting multiple engagement signals	Feed ranking, Reels, Ads
Feature interaction modeling	DeepFM, DCN for cross-feature learning	Ads, Recommendations
Embedding-based retrieval	Two-tower models for candidate generation	All recommendation systems
Calibration	Predicted probabilities must match actual rates	Ad auction (bid optimization)
Position bias correction	Users click more on items shown first	Feed ranking training data
Label definition	What counts as positive/negative engagement	Feed, Reels, Marketplace
Multi-objective optimization	Balancing engagement, integrity, time well spent	Platform-level optimization
Approximate nearest neighbor	Fast embedding lookup for billions of items	FAISS (Meta's ANN library)
Online learning	Model updates from real-time engagement signals	Ads, trending content
Privacy-preserving ML	Post-ATT ad targeting without cross-app tracking	Ads targeting

Meta Behavioral Story Bank

Prepare 5-7 stories that can be adapted to any Meta behavioral question:

Story Template	Maps To	Key Element
"I shipped X in Y weeks, which improved Z by N%"	Move Fast, Impact	Speed + quantified outcome
"I proposed an unconventional approach that the team initially resisted"	Be Bold	Data-driven persuasion, willingness to take risk
"I noticed a fairness issue and raised it despite pressure to ship"	Build Social Value	Ethical decision-making, proactive identification
"I simplified a complex system from N components to M"	Move Fast	Reducing complexity, not just adding features
"I collaborated with team X to solve a problem neither could solve alone"	Be Open	Cross-team collaboration, shared ownership
"I invested time in infrastructure that paid off 6 months later"	Long-Term Impact	Strategic thinking beyond quarterly goals
"I made a mistake, here's what I learned and changed"	Growth mindset	Accountability, learning, process improvement

Meta Offer Calibration (ML Roles, 2024-2025)

Approximate total compensation ranges to calibrate expectations (US, Bay Area):

Level	Title	Base	Bonus	RSU (Annual)	Total Comp
E3	IC3 (SWE)	$130-160K	10%	$50-100K	$200-280K
E4	IC4 (SWE)	$160-190K	15%	$100-200K	$290-430K
E5	IC5 (Senior)	$190-230K	15%	$200-350K	$430-620K
E6	IC6 (Staff)	$230-280K	20%	$350-600K	$650-950K

:::note Compensation notes:

• RSUs vest over 4 years with annual refreshers
• ML roles sometimes receive a premium over general SWE
• Research Scientist roles may be leveled differently
• These ranges shift annually; verify with current data :::

Next Steps

After completing Meta-Style preparation:

• Google-Style Problems if also interviewing at Google (different emphasis on approach discussion)
• Startup-Style Problems if also interviewing at startups (end-to-end, practical)
• MLE Problems for comprehensive MLE preparation
• AI Engineer Problems if your role involves LLM/GenAI work