Company interview guide

Mercor Machine Learning Interview Guide (2026)

Short screens, then a 3-4 hour onsite of algorithm coding, system design and a timed hands-on build where AI tools are often allowed.

Mercor's machine learning interview loop has 6 reported stages (Intro or behavioral call, AI-led interview, Technical screen or assessment, Onsite: algorithm coding, Onsite: system design, Onsite: hands-on build); typical length: Not well reported; the onsite is one 3-4 hour block.

Roles covered: ML Engineer and AI Engineer. The question bank holds 12 questions reported from Mercor interviews, 2 of them free to solve in your browser.

Try a free question

Interview process length
Not well reported; the onsite is one 3-4 hour block
Stages
6
Reports linked to bank questions
12
Bank report dates
Mar 2026–Sep 2026
Questions in bank
12 (2 free)
Last reviewed
October 2026

How the Mercor interview loop works

Based on about ten candidate reports, mostly from 2026. The usual shape is a short intro or behavioral call, one technical step (a timed online assessment, a take-home, or a live ML discussion), then an in-person onsite of three rounds lasting three to four hours: algorithm coding, system design, and a hands-on build such as training a small GPT or building a candidate search service against an evaluation endpoint. Some candidates also did a 30-minute AI-led interview, either first or alongside the onsite invite. ML engineering has at least two tracks (marketplace and frontier data products) and design prompts follow the team. One candidate heard back the same evening and another two days later. Reports mention weekend scheduling, a no-show interviewer and five-day in-office work, which the job postings confirm.

The Mercor interview process, stage by stage

01Intro or behavioral call

About 20 min with a recruiter or the hiring manager; sometimes this chat comes after the technical step

A short conversation about you and why Mercor. One candidate had a 20-minute behavioral screen before the onsite; another had a hiring-manager chat of standard behavioral questions after the online assessment, and was rejected despite being told the background matched.

What it tests

  • A crisp self-introduction
  • The work you are proudest of and the most intense stretch you have worked
  • Why Mercor, and how your background fits the team

How to prepare

  • Prepare a short story about high-intensity work; the company works in person five days a week and candidates say pace comes up.
  • Ask which team (marketplace or frontier data products) you are being considered for, since later design prompts follow the team.

Three candidate reports from 2025-2026 describe a short intro, behavioral or hiring-manager call; its place in the process varies.

02AI-led interview

About 30 min with an AI interviewer; in 2026 it was sent alongside the onsite invite, in 2024 it was the first step

Three candidates met an AI interviewer. The 2024 reports describe questions driven by the projects on the resume (plus SQL and Python), with few pauses between questions. A 2026 machine learning candidate was asked to design a social feed by an AI that kept pushing toward a preset answer. The 2024 reports may describe the screen for Mercor's talent marketplace rather than for a staff role.

What it tests

  • Explaining your own projects: models used, data, and day-to-day use
  • What you would improve in a past project and why
  • A system design prompt answered out loud, without a human to steer

How to prepare

  • Practise speaking about each resume project for two minutes, then answering 'what would you improve'.
  • Structure design answers explicitly (requirements, data model, serving path) so a scripted interviewer can follow them.
  • Treat it as a real round; it can sit alongside the human onsite.

Free reported questions for this stage

Two 2024 reports (ML engineer, ML researcher) on resume-driven AI interviews and one 2026 ML engineer report of an AI-led feed design round; few details.

03Technical screen or assessment

Varies: a 4-hour online assessment with AI coding tools allowed, a 72-hour open-ended take-home, or a 30-min live ML discussion

The technical step differs by team and over time. A 2026 applied AI candidate built an evaluation harness for a locally served LLM in a 4-hour assessment where AI coding tools were allowed. A 2025 applied AI candidate got a 72-hour take-home on optimising a batching system, with no stated objective. Other ML candidates had a live talk: one on post-training data selection and ablations, one on classic ML fundamentals.

What it tests

  • Building a small, working LLM tool end to end: dataset in, model calls, metrics, retries and failure logging
  • Choosing your own objective and trade-offs in an open-ended performance task
  • Judging whether noisy, mixed-source data helps post-training, and proving it with ablations
  • Core ML fundamentals

How to prepare

  • If AI tools are allowed, have a paid coding assistant ready beforehand, as one candidate advises.
  • For an open-ended take-home, state what you optimised for and measure it; one candidate was rejected without being told what to optimise.
  • For the data discussion, speak from real experience with filtering and ablations; a candidate said invented answers are easy to spot.

Four 2025-2026 reports, each a different format: a 4-hour assessment, a 72-hour take-home, a post-training data talk and an ML fundamentals call.

04Onsite: algorithm coding

30-45 min live, or 90 min with a 60-min recorded solo part (no LLMs, web search allowed) then 30 min live

The most consistent onsite round. Two 2026 candidates got the same problem: find the top k of n items using a comparator that is wrong about 10% of the time but always gives the same answer for the same pair, with follow-ups on minimising calls and on very slow calls. Others got a multi-part referral graph (reach, flow centrality, expected growth), exploring a friend network through a limited API, or a parking-spot puzzle.

What it tests

  • Robust selection and sorting when comparisons can be wrong but repeatable
  • Cutting the number of expensive calls, then cutting wall-clock time with parallel calls
  • Graph modelling with ranking and reachability, built up over several parts
  • Simple probabilistic estimates such as expected network size after N days

How to prepare

  • Practise top-k with a faulty comparator: tournaments, repeated or majority comparisons, and why a deterministic error defeats plain retries.
  • In the recorded solo format, leave clean, readable code; the interviewer reviews your first two parts before the third.
  • Count your calls aloud when an API or comparator is the costly resource.

Algorithm coding at the onsite in four 2026 ML engineer reports; the faulty-comparator top-k problem appears in two of them.

05Onsite: system design

About 45 min, discussion; prompt depends on the team

Prompts track the team. Frontier data candidates designed workflow orchestration for ML data pipelines (twice in our reports); a marketplace candidate designed a high-demand ticketing system; another designed candidate search from a job description. Expect general distributed-systems depth as well as ML, and pitch your answer at the interviewer's background.

What it tests

  • DAG workflows: state, retries, leases, idempotency and lineage
  • Inventory holds and preventing double-selling under traffic spikes
  • Retrieval plus ranking for candidate-job matching with hard filters
  • Failure handling and observability

How to prepare

  • Prepare one durable workflow-orchestration design in depth; it is the most repeated design prompt here.
  • A candidate stressed that communication is weighted heavily: check the interviewer follows before going deep on a niche area.

System design at the onsite in three 2026 ML engineer reports; workflow orchestration in two, candidate search and ticketing once each.

06Onsite: hands-on build

60-90 min: build alone for most of it (AI coding tools allowed in two of three reports), then 15-30 min of review and questions

A timed mini-project that must run end to end. Two candidates had to implement and train a small autoregressive language model (nanoGPT style, on a Shakespeare text) and then answer transformer questions. Another built a candidate search service in 75 minutes, scored it against an evaluation endpoint and iterated; the interviewer then asked how LLMs were used and how problems were found.

What it tests

  • Getting a decoder-only transformer to train and generate within the hour
  • Transformer concepts: pre- vs post-norm, self-attention, positional embeddings, decoding (greedy, beam, top-k, top-p)
  • Retrieval and ranking with LLM-generated filters for hard criteria
  • Iterating on measured results and explaining your debugging

How to prepare

  • Be able to write a minimal GPT and its training loop from memory, then check shapes, mask and loss shift quickly.
  • Use the AI tools to move fast, but be ready to defend every design choice; the review asks how you used them.
  • Submit something that runs end to end early, then improve the score.

Free reported questions for this stage

Three 2026 ML engineer reports: two trained a small GPT then took transformer questions; one built candidate search against an evaluation endpoint.

Questions and themes reported most often

  • Top k with a comparator that is wrong 10% of the time, deterministically (2 reports, onsite: algorithm coding)
  • Implement and train a small GPT end to end, then transformer and decoding questions (2 reports, onsite: hands-on build)
  • Candidate search from a job description with hard criteria (2 reports, onsite: hands-on build)
  • Design workflow orchestration for multi-step data pipelines (2 reports, onsite: system design)
  • Multi-part graph coding on a referral or friend network (2 reports, onsite: algorithm coding)
  • AI interviewer probing the projects on your resume (2 reports, ai-led interview)
  • Self-introduction, proudest work, most intense work, why Mercor (2 reports, intro or behavioral call)
  • Evaluation harness for a locally served LLM (1 reports, technical screen or assessment)

What Mercor asks about

  • System design (ML and general)27%
  • Coding & algorithms25%
  • Hands-on build with AI tools20%
  • LLM data, evaluation & concepts18%
  • Behavioral & resume10%

Mercor interview prep checklist

  • Solved top k with a faulty deterministic comparator and explained the call count
  • Trained a small GPT from scratch end to end in under an hour
  • Can explain pre- vs post-norm, positional embeddings and top-k vs top-p decoding
  • Sketched a workflow orchestrator with retries, leases and idempotent steps
  • Built a retrieve-then-rank search with hard filters and measured it
  • Rehearsed a two-minute walk-through of every resume project
  • Asked which team and which tool rules apply to each round

Tips

  • Expect evidence here to be thin: about ten reports, and formats changed between candidates, so ask the recruiter for the type, length and tool rules of every round.
  • Drill top-k selection with an unreliable but deterministic comparator, including how to cut calls and wall-clock time.
  • Be able to write and train a minimal GPT in under an hour, and revise norms, attention, positions and decoding strategies.
  • Prepare a durable workflow-orchestration design and a candidate search and ranking design; both map to the teams that are hiring.
  • Practise building with AI coding tools under a clock, and separately without them, since rules differ by round.
  • Have a real story about filtering noisy training data and proving it helped.
  • Plan for in-person work five days a week and possible weekend interview slots, both mentioned by candidates.
  • Keep explanations matched to your interviewer; a candidate said communication weighed heavily.

See the rest of the Mercor loop

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

Unlock with Pro

Practice the Mercor loop

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

Mercor interview FAQ

How many interview rounds does Mercor have for ML roles?

Candidates report 6 stages: Intro or behavioral call, AI-led interview, Technical screen or assessment, Onsite: algorithm coding, Onsite: system design and Onsite: hands-on build. Based on about ten candidate reports, mostly from 2026. The usual shape is a short intro or behavioral call, one technical step (a timed online assessment, a take-home, or a live ML discussion), then an in-person onsite of three rounds lasting three to four hours: algorithm coding, system design, and a hands-on build such as training a small GPT or building a candidate search service against an evaluation endpoint. Some candidates also did a 30-minute AI-led interview, either first or alongside the onsite invite. ML engineering has at least two tracks (marketplace and frontier data products) and design prompts follow the team. One candidate heard back the same evening and another two days later. Reports mention weekend scheduling, a no-show interviewer and five-day in-office work, which the job postings confirm.

How long does the Mercor ML interview process take?

Reported timelines run Not well reported; the onsite is one 3-4 hour block.

What does Mercor ask in ML interviews?

By reported question mix: System design (ML and general) 27%; Coding & algorithms 25%; Hands-on build with AI tools 20%; LLM data, evaluation & concepts 18%; Behavioral & resume 10%. The most repeated themes: Top k with a comparator that is wrong 10% of the time, deterministically; Implement and train a small GPT end to end, then transformer and decoding questions; Candidate search from a job description with hard criteria.

How should I prepare for the Mercor ML interview?

Expect evidence here to be thin: about ten reports, and formats changed between candidates, so ask the recruiter for the type, length and tool rules of every round. Drill top-k selection with an unreliable but deterministic comparator, including how to cut calls and wall-clock time. Be able to write and train a minimal GPT in under an hour, and revise norms, attention, positions and decoding strategies.

Are there free Mercor interview questions?

Yes. LeetML has 12 questions reported from Mercor interviews, and 2 of them are free to solve in the browser; they are listed by stage above.

Related interview guides