Company interview guide

Scale AI Machine Learning Interview Guide (2026)

Hands-on notebook work on messy data and buggy ML code, LLM coding from scratch, and a values-driven behavioral loop.

Scale AI's machine learning interview loop has 8 reported stages, from recruiter and hiring-manager calls to credo, customer engagement and hiring manager; typical length: Several weeks.

Roles covered: ML Engineer, AI Engineer and Research Scientist. The question bank holds 32 questions reported from Scale AI interviews, 3 of them free to solve in your browser.

Try a free question

Meta owns 49%. Meta bought a 49% non-voting stake in June 2025, and Scale's founder and CEO left to lead Meta's superintelligence lab. Scale remains an independent company and runs its own hiring.

Interview process length
Several weeks
Stages
8
Reports linked to bank questions
37
Bank report dates
Jan 2021–Sep 2026
Questions in bank
32 (3 free)
Last reviewed
October 2026

How the Scale AI interview loop works

A recruiter call, often a hiring-manager or senior-leader call, then a technical screen; in older ML loops and some recent ones a take-home comes first. The final loop is usually four interviews, sometimes split over several days: coding, ML or LLM fundamentals, system design or an applied build, a project presentation and a values (credo) behavioral round. Several candidates describe slow or missing follow-up between stages, so keep your recruiter informed and ask for dates in writing.

The Scale AI interview process, stage by stage

01Recruiter and hiring-manager calls

A short recruiter call, often plus a hiring-manager or senior-leader call before or after the technical screen

A background and motivation check. In many loops a manager or VP call follows, sometimes before the technical screen and sometimes after it; one 2026 manager call focused on the candidate's current project, and one 2025 AI engineer loop had a second manager call at the end.

What it tests

  • Why Scale, and how your work relates to data, evaluation or post-training
  • A clear, metric-backed summary of your current project
  • Logistics: level, location, in-office expectations, timeline

How to prepare

  • Ask which track (CV, NLP, LLM, applied, infra) your later rounds will follow; interviewers have been mismatched to candidates before.
  • Confirm whether there is a take-home and its deadline before you schedule anything else.

Recruiter or HR call in 9 distinct reports from 2019 to 2026 and a manager or VP call in 5; the order of the manager call relative to the screen varies.

02Take-home or online assessment

Classic version: 1-2 weeks, choose a CV or NLP project; one 2025 variant was a 6-hour paper-implementation notebook

The long-running version gives you a dataset and a project skeleton and asks you to train a model past an accuracy bar, with a penalty above about 5M parameters; the CV option is usually detecting rotated five-point stars with rotated boxes. Some candidates were rejected on the take-home alone. Since 2025 it appears in only a few reports, including a GitHub-repo assignment and a six-hour notebook in which you implement an adversarial-prompt paper against a small open model.

What it tests

  • Training a working model from a skeleton, with parts you wrote yourself
  • Choosing a loss and output head for an unusual target such as box angle
  • Staying under a model-size cap while hitting an accuracy threshold
  • Turning a research paper into running code

How to prepare

  • Write up your loss and design choices; one interviewer asked about them, though other candidates said later rounds never mentioned the project.
  • Handle angle periodicity and symmetry explicitly for rotated-box targets.
  • Budget a full day or more even if told it takes a few hours, and pin your environment early; one candidate lost time to package conflicts in the provided notebook.

A take-home or assessment appears in 29 distinct reports, 25 of them from 2020-2024 (mostly the CV or NLP project); only 4 reports since 2025 mention one.

03Technical screens

40-60 min live in a Colab or Jupyter notebook (open book in some reports); some loops add a 30-min LLM fundamentals screen, and applied AI engineer loops use object-oriented coding

The most consistent round across years: load raw images, text or JSON, shape it for a provided model, run batch inference, compute metrics and list misclassified items, then often debug three planted bugs. Screens from 2020-2022 added a numpy simulation of an order-statistic gap. Recent variants include an LLM and post-training fundamentals screen and, for applied AI engineers, a 60-minute card-game coding exercise.

What it tests

  • Fast, correct data wrangling with numpy and pandas: files, encodings, shapes
  • Batch inference and evaluation of models you did not write
  • Locating planted bugs (off-by-one, leakage, loss reduction, shapes, seeding) under time pressure
  • Rapid-fire LLM basics: attention cost, causal masks, decoding, RLHF, evaluation
  • Simulating a distribution, plotting it and naming it

How to prepare

  • Use tracebacks and a tiny batch to find bugs instead of reading the whole repository.
  • Write the data loader defensively: count and log malformed lines rather than crashing.
  • Keep LLM fundamentals answers short and concrete; one candidate felt they missed what the interviewer wanted by going broad.
  • Expect a time crunch; finish the core path end to end before polishing.

Free reported questions for this stage

A live technical screen in 34 distinct reports; about 12 describe notebook data handling and inference, 5 planted-bug debugging and 5 the order-statistic simulation.

04Final loop: coding

About 60 min each; LLM coding in Python or numpy, a general data-structures problem, or an object-oriented exercise

Research and LLM-track loops ask you to implement multi-head causal self-attention and sampling methods such as top-p from scratch. Other loops include a general coding problem, such as building a graph from a dictionary and finding a shortest path, and applied AI engineer loops favour an object-oriented card-game exercise whose rules are shared ahead of time. One 2026 infra-role loop had an AI-assisted practical: a small HTTP service that turns a CSV into JSON with an LLM-generated field.

What it tests

  • Attention shapes, head splitting and causal masking without a framework
  • Top-p, top-k and temperature sampling with correct edge cases
  • Graph and interval problems framed around annotation work
  • Clean class design for rules that grow, such as wildcards
  • Using an AI assistant productively while still owning the design

How to prepare

  • Be able to write masked multi-head attention and nucleus filtering in under 15 minutes each.
  • For the card exercise, model ranks and suits first, then add jokers as a search over substitutions.
  • In AI-enabled rounds, write your own prompts; pasting the task straight in has been stopped by interviewers.

Free reported questions for this stage

Final-loop coding in 11 distinct reports: 3 LLM attention or sampling (2025-2026), 4 general coding, 2 object-oriented, plus research-method coding; formats differ by track.

05Final loop: ML & LLM fundamentals

About 60 min discussion; some versions hand you a research paper to read and explain

A conceptual round that has covered overfitting and regularization, CNNs versus vision transformers, tokenizers, and reading a research paper to explain its inputs, outputs and key ideas. One applied AI engineer was rejected after this round and felt the interviewer wanted enthusiasm for LLM research. Post-training topics such as SFT, preference methods, reward hacking and verifiable rewards come mostly from public question lists and one 2026 screen, so treat them as likely themes.

What it tests

  • Explaining a paper's contribution quickly and accurately
  • Post-training choices: SFT, RLHF, DPO and verifiable rewards, and what data each needs
  • Failure modes such as reward hacking and how to detect them
  • Architecture trade-offs: decoder-only versus encoders, multimodal alignment

How to prepare

  • Practise summarising a paper you have not read in 15 minutes: problem, method, inputs, outputs, results.
  • Tie post-training answers to data, since data is Scale's business; say what labels you need and how you would check their quality.

An ML or research fundamentals round in 6 distinct reports from 2023 to 2026; 3 of them include reading and explaining a paper. Post-training items rest on public question lists.

06Final loop: system design and applied ML

About 60 min with a shared doc or whiteboard; applied AI engineer loops have an applied ML coding round instead

Few ML-role reports describe this round in detail. An AI engineer loop listed system design, an applied AI engineer loop coded a simple RAG pipeline, and an infra-role candidate in an ML research org designed an embedding and classification API with model versions, batch and synchronous paths, and tenant isolation, saying all AI roles get it. Data-quality and evaluation designs fit Scale's products but come from public question lists.

What it tests

  • API and version contracts for model-serving systems
  • Batching, accelerator use and multi-tenant isolation
  • Measuring label quality and annotator agreement
  • Designing evaluations and debugging a sudden score change
  • Retrieval pipelines with access control

How to prepare

  • Version the whole inference path (preprocessing, embedding model, classifier) and reject incompatible pairs.
  • For any data or eval design, state the quality metric and the audit loop before the architecture.
  • Keep it metrics-driven; one recent interviewer was a general software engineer, not a domain specialist.

System design or applied ML in 3 ML-role reports (2019, 2025, 2026); the detailed API design comes from one infra-role report, and eval designs from public question lists.

07Technical deep dive or presentation

About 45-60 min presentation of a past project, with questions

You present a project of your own, sometimes with slides. In the one detailed account the audience was a general software engineer, so explain domain ideas simply and anchor every claim in numbers.

What it tests

  • Clear problem framing and your personal contribution
  • Metrics, baselines and what moved them
  • Trade-offs and what you would change

How to prepare

  • Prepare slides that a general engineer can follow, with one results slide that carries the story.
  • Rehearse answers on scale, failure cases and why you chose your approach.

A 45-min presentation in one 2026 MLE report; a 2026 infra-role report and a public interview guide describe the same round.

08Credo, customer engagement and hiring manager

1-2 conversations of 30-60 min, including a values (credo) round and in 2026 a customer-engagement round

Behavioral rounds map to Scale's published credo values, such as earning customer love and treating quality as an advantage. Common prompts are your most meaningful or hardest project, motivation, weaknesses and working without direction. One 2026 candidate believed this was where they failed, after unusual questions about their current manager. One interviewer mix was all CV for an NLP candidate, so ask about the panel.

What it tests

  • Working with demanding customers or stakeholders and driving an outcome
  • Ownership on projects with little direction
  • Honest self-assessment: strengths, weaknesses, how others rate you
  • Focus on the few things that matter most

How to prepare

  • Read the credo values on the careers page and prepare one story for each.
  • Have a calm, factual answer ready for how your current manager would rate you and why.
  • Bring questions about team culture and office expectations; one candidate got a candid, useful answer this way.

A behavioral or culture round in 10 distinct reports from 2021 to 2026 (one from an infra role); the credo values are listed on the careers page, which does not describe the round.

Questions and themes reported most often

  • Take-home: detect rotated stars (CV) or an NLP task under a model-size cap (15 reports, take-home or online assessment)
  • Notebook data wrangling: load raw files or images, batch inference with given models, compute metrics (12 reports, technical screens)
  • Behavioral and credo questions, including handling an opinionated customer (6 reports, credo, customer engagement and hiring manager)
  • Prepare the data, then find three planted bugs in an ML or LLM project (5 reports, technical screens)
  • Gap between the 5th and 6th of ten uniform draws: simulate, plot, name the distribution (5 reports, technical screens)
  • Read a research paper, then explain or implement it (5 reports, final loop: ml & llm fundamentals)
  • Rapid-fire LLM basics: attention cost, masks, decoding, tokenizers, RLHF, evaluation (4 reports, technical screens)
  • General coding, e.g. build a graph from a dictionary and find a shortest path (4 reports, final loop: coding)

What Scale AI asks about

  • ML coding & debugging35%
  • ML & LLM concepts20%
  • General & OOP coding15%
  • ML / LLM system design & evals12%
  • Behavioral & deep dive10%
  • Probability & statistics8%

Scale AI interview prep checklist

  • Completed a timed notebook drill: load raw data, batch inference, metrics and error analysis in under an hour
  • Found and fixed three planted bugs in an unfamiliar training script
  • Wrote masked multi-head attention and nucleus sampling from scratch
  • Can explain SFT, RLHF, DPO and verifiable rewards, with the data each needs
  • Sketched a versioned, multi-tenant embedding and classification service
  • Prepared a metrics-driven project presentation for a non-specialist audience
  • Have a story for each credo value, including a difficult customer

Tips

  • Drill the notebook workflow until it is automatic: read messy files, reshape to batches, run inference, compute metrics, list the misclassified items.
  • Practise finding planted bugs in unfamiliar PyTorch code from tracebacks and a tiny batch; leakage, loss reduction and off-by-one tokenization are common suspects.
  • Write masked multi-head attention and top-p sampling from memory in numpy.
  • Know your post-training basics cold: SFT, RLHF with KL control, DPO, verifiable rewards and reward hacking.
  • Connect answers to data quality; Scale sells data and evaluation, so say how you would measure label and benchmark quality.
  • If you get a take-home, document design decisions; one interviewer quizzed a candidate on the loss choice.
  • Prepare one credo story per value and one metrics-driven project presentation.
  • Follow up politely but firmly; several candidates report long silences between stages.

See the rest of the Scale AI loop

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

Unlock with Pro

Practice the Scale AI loop

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

Scale AI interview FAQ

How many interview rounds does Scale AI have for ML roles?

Candidates report 8 stages: Recruiter and hiring-manager calls, Take-home or online assessment, Technical screens, Final loop: coding, Final loop: ML & LLM fundamentals, Final loop: system design and applied ML, Technical deep dive or presentation and Credo, customer engagement and hiring manager. A recruiter call, often a hiring-manager or senior-leader call, then a technical screen; in older ML loops and some recent ones a take-home comes first. The final loop is usually four interviews, sometimes split over several days: coding, ML or LLM fundamentals, system design or an applied build, a project presentation and a values (credo) behavioral round. Several candidates describe slow or missing follow-up between stages, so keep your recruiter informed and ask for dates in writing.

How long does the Scale AI ML interview process take?

Reported timelines run Several weeks (Several weeks; the take-home alone allowed 1-2 weeks).

What does Scale AI ask in ML interviews?

By reported question mix: ML coding & debugging 35%; ML & LLM concepts 20%; General & OOP coding 15%; ML / LLM system design & evals 12%; Behavioral & deep dive 10%; Probability & statistics 8%. The most repeated themes: Take-home: detect rotated stars (CV) or an NLP task under a model-size cap; Notebook data wrangling: load raw files or images, batch inference with given models, compute metrics; Behavioral and credo questions, including handling an opinionated customer.

How should I prepare for the Scale AI ML interview?

Drill the notebook workflow until it is automatic: read messy files, reshape to batches, run inference, compute metrics, list the misclassified items. Practise finding planted bugs in unfamiliar PyTorch code from tracebacks and a tiny batch; leakage, loss reduction and off-by-one tokenization are common suspects. Write masked multi-head attention and top-p sampling from memory in numpy.

Are there free Scale AI interview questions?

Yes. LeetML has 32 questions reported from Scale AI interviews, and 3 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources