Meta

English · Русский · AI Interview Atlas › Companies

Big Tech · International · Loop reviewed 2026-09-26

Structured loops with published preparation guides. Recent reports describe an AI-assisted coding round for engineering managers and live prototyping in AI product manager interviews.

On this page: Interview loop · Official preparation material · Questions (20)

Roles covered

• AI Leadership: Engineering manager, AI product manager (senior to group), Technical program manager

Interview loop

Stage	What happens	Basis
Engineering manager initial interview	A 45-minute conversation with a software engineering manager covering both technical and management skills.	✅ confirmed by the company Preparing for Your Initial Interview (software engineering managers), Meta Careers, retrieved 2026-09-26
Technical program manager loop	Five conversations with technical program managers and cross-functional team members.	✅ confirmed by the company Preparing for Your Interviews at Meta (technical program managers), Meta Careers, retrieved 2026-09-26
Engineering manager report (M1)	Screening with a behavioral round and a system design round chosen from ML, software or product; the main loop had ML system design, a project deep dive, team management, behavioral questions and an AI-assisted coding round.	🗣 candidate report Meta Engineering Manager (M1) Interview Experience, Aced (formerly Exponent), candidate report, retrieved 2026-09-26
AI product manager report (level unverified)	Product sense and analytical phone screens, then an hour split between product sense and live prototyping in a Llama-based vibe-coding tool, with follow-ups on retrieval, latency, inference cost and token efficiency.	🗣 candidate report Meta Product Manager, AI Interview Experience (level unverified), Aced (formerly Exponent), candidate report, retrieved 2026-09-26

Coding: varies by role. Engineering managers may get an AI-assisted coding round (Meta Engineering Manager (M1) Interview Experience, Aced (formerly Exponent), candidate report, retrieved 2026-09-26)

Official preparation material

• Preparing for Your Initial Interview (software engineering managers), Meta Careers, retrieved 2026-09-26
• Preparing for Your Interviews at Meta (technical program managers), Meta Careers, retrieved 2026-09-26

Questions reported for Meta

LLM fundamentals

• Compare multi-head, multi-query and grouped-query attention in quality and serving memory.
  • Knowledge · Asked at: Meta † · ✍ Answer
  • Tests: Whether shared keys and values are distinguished from query heads.
  • A strong answer covers:
    • MHA uses separate KV heads; MQA shares one; GQA shares within groups.
    • Relate KV-head count to cache size and memory reads.
    • Discuss quality evaluation and the need to train or adapt the checkpoint.
  • Read: GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints (Ainslie et al., arXiv)
• Explain rotary position encoding and the trade-offs of interpolation when extending context beyond training lengths.
  • Knowledge · Asked at: Meta †
  • Tests: Whether a longer accepted input is distinguished from reliable long-context reasoning.
  • A strong answer covers:
    • Explain RoPE: rotate query and key dimension pairs by position times a per-pair frequency, so scores depend on offset.
    • Show that position interpolation divides every frequency by the scale, blurring nearby tokens; NTK and YaRN mostly rescale low frequencies.
    • Separate accepted length from use: test retrieval and multi-hop tasks at varied depths, plus short-context regressions after finetuning.
  • Read: RoFormer: Enhanced Transformer with Rotary Position Embedding (Su et al., arXiv) · Lost in the Middle: How Language Models Use Long Contexts (Liu et al., arXiv) · Extending Context Window of Large Language Models via Positional Interpolation (Chen et al., arXiv) · YaRN: Efficient Context Window Extension of Large Language Models (Peng et al., arXiv) · RULER: What's the Real Context Size of Your Long-Context Language Models? (Hsieh et al., arXiv)
• What objectives and data distinguish pretraining, instruction tuning and preference optimisation?
  • Knowledge · Asked at: Meta †
  • Tests: Whether the candidate can identify what each stage can and cannot change.
  • A strong answer covers:
    • Name the objectives: next-token loss on raw corpora, the same loss on demonstrations masked to responses, then pairwise preferences.
    • State what each changes: pretraining sets knowledge and skills, instruction tuning sets format, preferences re-rank behaviours already available.
    • Probe the limits: facts added only in tuning generalise poorly, and preference training can reward length or flattery.
  • Read: Training language models to follow instructions with human feedback (Ouyang et al., arXiv) · Direct Preference Optimization: Your Language Model is Secretly a Reward Model (Rafailov et al., arXiv) · Does Fine-Tuning LLMs on New Knowledge Encourage Hallucinations? (Gekhman et al., arXiv)
• Explain the two branches of SwiGLU and the parameter and compute trade-offs against a conventional feed-forward block.
  • Knowledge · Asked at: Meta †
  • Tests: Whether gated activations are explained beyond their name.
  • A strong answer covers:
    • Write SwiGLU: Swish of one projection gates a second linear projection elementwise, then a down projection returns to d_model.
    • Count parameters: three matrices instead of two, so hidden width shrinks to about 8d/3 to keep 8d² weights.
    • Compare at matched parameters and FLOPs on held-out loss, noting extra activation memory and fused gate-up kernels.
  • Read: GLU Variants Improve Transformer (Shazeer, arXiv)

Inference, serving and cost

• Compare tensor, pipeline, data, sequence and expert parallelism for a large model deployment.
  • Knowledge · Asked at: Amazon †, Google and Google DeepMind †, Meta †
  • Tests: Whether communication patterns and memory limits determine the choice.
  • A strong answer covers:
    • Name each split and its traffic: TP per-layer all-reduce, PP stage activations, SP key-value exchange, EP token all-to-all.
    • Fit weights and KV first: TP inside the NVLink domain, PP across nodes, then traffic-free data-parallel replicas for throughput.
    • Measure per-GPU memory headroom, communication share of step time, pipeline bubbles and expert load imbalance at target latency.
  • Read: vLLM documentation (vLLM project) · Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity (Fedus et al., arXiv) · Efficiently Scaling Transformer Inference (Pope et al., arXiv) · Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism (Shoeybi et al., arXiv)

Safety, security and governance

• Design a system for detecting harmful content in Facebook and Instagram uploads.
  • System design · Senior · Asked at: Meta †
  • Tests: Whether you combine detection, review and appeals at the required scale.
  • A strong answer covers:
    • Tier harms by severity and choose per tier: block at upload, limit distribution or queue review.
    • Combine hash matching for known content with classifiers for new content, routing uncertain cases to reviewers.
    • Measure prevalence, per-policy precision and appeal overturn rate, weighing wrongful removals against missed harm.
  • Read: A Holistic Approach to Undesired Content Detection in the Real World (Markov et al., arXiv) · Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations (Inan et al., arXiv)

Multimodal and voice

• Compare projection layers, cross-attention and token-based integration for giving a language model access to images.
  • Knowledge · Asked at: Meta † · ✍ Answer
  • Tests: Whether image encoding is separated from cross-modal fusion.
  • A strong answer covers:
    • Separate the vision encoder from the interface to the language model.
    • Compare projected patch embeddings, cross-attention to visual features and discrete image tokens.
    • Discuss visual token budget, spatial detail and training alignment.
  • Read: Flamingo: a Visual Language Model for Few-Shot Learning (Alayrac et al., arXiv)
• What architectural and data choices change when a vision-language model must reason over video rather than isolated images?
  • Knowledge · Asked at: Meta †
  • Tests: Whether temporal information and sampling cost are addressed.
  • A strong answer covers:
    • Budget visual tokens: frames times tokens per frame explode, so choose frame rate, pooling or tubelet patches per task.
    • Encode time explicitly with timestamps or temporal positions, and train on clips whose answers depend on event order.
    • Test against single-frame and shuffled-frame baselines to confirm the model uses temporal information at all.
  • Read: Flamingo: a Visual Language Model for Few-Shot Learning (Alayrac et al., arXiv)

AI system design

• Design a moderation system that combines specialist classifiers and LLMs while controlling delay and review workload.
  • System design · Asked at: Meta † · ✍ Answer
  • Tests: Whether risk classes, escalation and false decisions shape system design.
  • A strong answer covers:
    • Define risk classes by harm, with separate tolerance for wrongful removal and missed violations per class.
    • Cascade: fast classifiers on everything, an LLM on the uncertain band, people on severe ambiguous cases.
    • Set thresholds per class from measured precision and recall, tracking reviewer queue, appeals and overturns.
  • Read: A Holistic Approach to Undesired Content Detection in the Real World (Markov et al., arXiv) · Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations (Inan et al., arXiv)
• Design a streaming chat service for hundreds of millions of users, including capacity, conversation storage and graceful overload behaviour.
  • System design · Asked at: Anthropic †, Google and Google DeepMind †, Meta †, OpenAI † · ✍ Answer
  • Tests: Whether scale assumptions lead to explicit capacity and reliability decisions.
  • A strong answer covers:
    • Estimate concurrent streams, token demand and storage from declared traffic assumptions.
    • Separate admission control, inference scheduling and durable conversation state.
    • Plan capacity headroom, cancellation, overload rejection and recovery from partial failure.
  • Read: Site Reliability Engineering (Google) · Efficient Memory Management for Large Language Model Serving with PagedAttention (Kwon et al., arXiv)

AI product strategy and metrics

• Notification engagement rises while total time spent remains flat. How would you interpret that result?
  • Product case · Senior · Asked at: Meta 🗣 · ✍ Answer
  • Tests: Whether you distinguish substitution, measurement artifacts and real user value.
  • A strong answer covers:
    • Check definitions, exposure, cohorts and the experiment setup.
    • Test whether notifications shift existing activity or create valuable new actions.
    • Include opt-outs, satisfaction and longer-term effects before deciding.
  • Read: Measuring the User Experience on a Large Scale: User-Centered Metrics for Web Applications (Rodden, Hutchinson and Fu, Google Research)
• Design a product connecting local tradespeople with customers.
  • Product case · Senior · Asked at: Meta 🗣
  • Tests: Whether you reason about both sides of a marketplace and initial liquidity.
  • A strong answer covers:
    • Name both sides, the decision each makes, and what each must trust: vetting, price, payment.
    • Measure requests ending in a completed job and time to match, guarded by disputes and repeat bookings.
    • Launch in one area and one trade, recruiting tradespeople by hand before automating the match.
  • Read: How to match supply and demand in your marketplace (Sharetribe)
• A new service grows while Facebook monthly active users decline. How would you assess whether the service is succeeding?
  • Product case · Senior · Asked at: Meta 🗣
  • Tests: Whether you investigate portfolio effects instead of reading one metric in isolation.
  • A strong answer covers:
    • Ask whether the service should add new users and time or move existing time between apps.
    • Estimate substitution with a holdout or matched cohorts, not by comparing the two curves.
    • Judge it on net portfolio users and value, with the core product's health as a guardrail.
  • Read: Patterns of Trustworthy Experimentation: Pre-Experiment Stage (Microsoft Research) · Measuring the User Experience on a Large Scale: User-Centered Metrics for Web Applications (Rodden, Hutchinson and Fu, Google Research)
• Design a volunteer-matching product and explain how it gets through cold start.
  • Product case · Senior · Asked at: Meta 🗣
  • Tests: Whether you make the first useful matches possible before scale arrives.
  • A strong answer covers:
    • Pick which side is scarce, then narrow to one city and cause where matches can happen.
    • Measure posted needs filled by a completed shift, with no-shows and organiser churn as guardrails.
    • Match the first organisations and volunteers by hand to test whether matches turn into shifts.
  • Read: How to match supply and demand in your marketplace (Sharetribe)

Leading engineering teams

• How have you managed and developed your strongest performers?
  • Behavioral · Senior · Asked at: Meta 🗣
  • Tests: Whether you provide challenge and growth without creating a privileged exception.
  • A strong answer covers:
    • Start from what each strong performer wants next, and stretch them with harder problems, not more volume.
    • Avoid a privileged exception: same standards on collaboration, and stretch work offered openly, not by favour.
    • Close with evidence of their growth, such as scope they now own alone, and whether they stayed.
  • Read: [re:Work] Google's New Manager Training Slides (Google re:Work)
• A cross-functional partner gives you critical feedback about your leadership. How do you respond?
  • Applied scenario · Senior · Asked at: Meta 🗣
  • Tests: Whether you investigate feedback and translate it into changed behaviour.
  • A strong answer covers:
    • Ask for specific instances and their impact on the partner's team, holding back any defence at first.
    • Test whether it is a pattern with other partners and your team, separating your behaviour from structural friction.
    • Commit to one visible change, tell the partner what it is, and ask them later whether it worked.
  • Read: [re:Work] Google's New Manager Training Slides (Google re:Work)
• What have you changed about your management approach as you have grown?
  • Behavioral · Senior · Asked at: Meta 🗣
  • Tests: Whether reflection produces concrete changes in how you lead.
  • A strong answer covers:
    • Name one or two specific practices you changed, not a general shift in philosophy.
    • Explain what triggered each change, such as feedback or a team outcome, and what you believed before.
    • Show the effect as your team would describe it, and name what you are still changing.
  • Read: [re:Work] Google's New Manager Training Slides (Google re:Work)

Programmes and delivery

• Describe a large cross-functional project, including ambiguity, roadblocks and delays.
  • Behavioral · Senior · Asked at: Meta 🗣 · ✍ Answer
  • Tests: Whether your delivery method adapts to changing information and dependencies.
  • A strong answer covers:
    • Map deliverables and dependencies to accountable owners.
    • Explain how you detected a blockage and changed the plan or escalation path.
    • Distinguish actual delivery and impact from initial targets.
  • Read: Design Docs at Google (Industrial Empathy) · Site Reliability Engineering (Google)

Behavioral and values

• Describe a project that used data and machine learning. What obstacles did you encounter?
  • Behavioral · Senior · Asked at: Meta †
  • Tests: Whether you connect technical obstacles to actions and delivered outcomes.
  • A strong answer covers:
    • Pick obstacles specific to data and models — labels, leakage, drift, evaluation gaps — not generic delays.
    • For each obstacle, say how you diagnosed it and why you chose that fix over the alternatives.
    • End with the delivered outcome measured against a baseline, and which obstacle you would now catch earlier.
  • Read: Rules of Machine Learning (Google for Developers)
• Tell me about achieving a meaningful result despite substantial ambiguity.
  • Behavioral · Senior · Asked at: Meta †
  • Tests: Whether you reduce uncertainty through decisions and feedback.
  • A strong answer covers:
    • Choose a case where the goal or the problem itself was unclear, not just resources or time.
    • Show which unknown you resolved first, with what small decision or probe, and how feedback changed the plan.
    • Close with the meaningful result and the evidence that you had solved the right problem.
  • Read: How the alpha phase works (Government Digital Service)

← Google and Google DeepMind · Microsoft →