Mercor Interview Process & Rounds
Mercor's software engineering process in 2025-2026 reports is a short background or hiring-manager call, a 30-minute technical phone screen that is entirely spoken, and an in-person onsite in San Francisco that runs three to five hours, sometimes followed by a short hiring-manager or executive chat. Decisions are fast: onsite results came the same day to two days later in the reports that give timing. The phone screen rewards answering four or five quick questions briefly rather than going deep on any one: one candidate who covered only three failed, and a commenter says you need at least four.
Part of the full catalog of Mercor interview questions, covering every round.
AceOffer is not affiliated with, endorsed by, or sponsored by Mercor. Questions are reconstructed from interview experiences candidates shared, not from Mercor materials.
Key facts
- •4 distinct round types
- •8 questions reconstructed from 22 candidate reports
- •Reports span Oct 2025 – Oct 2026
- •Refreshed monthly · last updated November 2026
The Mercor loop, from candidate reports
Mercor's loop opens with a ~30-minute technical phone screen that is entirely verbal: about four rapid-fire questions (sorting theory, a probability puzzle, code reviews of a threaded cache and a CSV writer, an idempotent payment call), answered as many as you can in the time, and two candidates report failing after answering three. It moves to an in-person onsite in San Francisco that runs several hours and mixes a multi-part coding round, a system-design round, an algorithm round, and a 90-minute candidate-search build (75 minutes writing it with LLM tools and running it against an evaluation endpoint, then 15 minutes of discussion); for MLE roles the onsite adds a one-hour build where you implement and train a small autoregressive language model (nanoGPT) on Shakespeare, then a transformer-fundamentals review. The algorithm round can be entirely spoken (one report: find an exchange rate across currency pairs). Applied AI candidates get a different screen: debug an async refund function, then rapid-fire questions on idempotency, RAG versus fine-tuning, context limits, statefulness and prompt injection. The signature onsite coding problem is a 3-part referral network (async, screen-recorded, no LLM for the first two parts; the third done live), and the signature system design is a distributed job scheduler / work orchestration (candidates are pointed at the HelloInterview job-scheduler design). One report stresses that interviewers weight communication fit heavily (tailor depth to the interviewer's domain), and single reports describe an interviewer no-show and contradictory instructions about whether LLMs are allowed in the coding round.
Mercor interview timeline
| Step | Typical timing |
|---|---|
| Phone screen to result | About one day (1 report) |
| Onsite to decision | Same day to two days (4 reports) |
Few reports give dates. One MLE candidate received an onsite invitation immediately after the phone screen, and says the default interview slots offered were on weekends. One candidate had a recruiter call and interview booked for the same day and was rejected by an early-morning email before either happened, because the position had been filled.
The Mercor interview process, step by step
1. Background or hiring-manager call
Short call before the technical screen; 20 minutes in one report
The first conversation varies. One candidate who applied online says there was no recruiter screen: the first call was with a hiring manager asking basic questions. Another describes a background round that probed one recent project from a product angle (was the decision yours or upstream, how did you define success, what were the metrics), then asked what Mercor does, why you want to join, and in detail how you use AI tools day to day. An Applied AI candidate had a 20-minute behavioral call: tell me about yourself, work you are proud of, your most intense work. One commenter says a survey before the phone screen asked about weekend availability.
- →Prepare a specific answer to why Mercor and what it does; one candidate says this is always asked in that round.
- →Be ready to describe concretely how you use AI tools, including which ones and for what (on-call, code review, design).
2. Technical phone screen
About 30 minutes, spoken answers only, no code written
Nine reports put the software engineering screen at about 30 minutes (one ran to 40) and one at about 50, and every report that describes the format says you answer out loud without writing code. It is rapid-fire: reports describe four questions, and one commenter heard there are five and that you need four. Recurring themes are sorting complexity (a faster-than-quadratic sort, then the slowest sort you can think of), a short probability puzzle, reviewing a small piece of code such as a cache used by several threads or a CSV writer, and a payment call where you discuss retries and duplicate charges. One candidate who answered only three questions failed, a commenter put another candidate's rejection down to the same thing, and one who answered four still did not pass. You can choose the language you discuss code in; one interviewer said they keep the same problem in twelve languages.
- →Keep each answer short and move on. The format rewards covering every question.
- →For code review, go beyond style: candidates were pushed toward race conditions, locking scope, cache eviction and input validation.
- →For the payment question, talk about idempotency keys, recording the payment before charging, and what database guarantees you need.
- →Pick a language whose threading model you can explain; one candidate who started in JavaScript switched to C++ to discuss locks.
3. Role-specific screens
20 to 45 minutes, varies by team
Other roles get different screens. An infrastructure candidate had 45 minutes and about 20 quick questions with no coding, spanning networking, processes and threads, streaming delivery guarantees and Kubernetes. An Applied AI candidate debugged an async refund function with retries, then answered quick questions on idempotency, RAG versus fine-tuning, context limits, making LLM apps stateful and prompt injection. A data engineering candidate expected quick questions and instead got a 20-minute case study on modeling external career data. A research engineering candidate had a conversation about post-training data selection and ablation studies, which they say you cannot fake without hands-on experience. An MLE candidate had 30 minutes of ML fundamentals, and was then asked, along with the onsite invitation, to complete a 30-minute chat with an AI interviewer that had them design a social feed; they describe it looping until it got the answer it expected. One Applied AI applicant instead started with a four-hour take-home building an evaluation harness with an Ollama integration, with AI coding tools allowed, then a hiring-manager chat, and was rejected after the manager had said their background matched.
- →Ask the recruiter whether the screen is Q&A, a case study or a debugging exercise; one candidate was caught off guard by a case study.
- →If you get a take-home that allows AI coding tools, the one candidate who reported it advises paying for the subscription before you start.
4. Onsite: coding round
In person in San Francisco; about 90 minutes: 60 minutes alone (screen recorded), then 30 minutes live review
The onsite is in person at the San Francisco office; one out-of-state candidate was told remote was not an option. Reported days run noon to 4pm and noon to 5pm, and an Applied AI loop was three hours. Task links arrive by email on the day, and several reports call the day disorganized, including a meeting room double-booked mid-interview and conflicting instructions on AI tools. For the coding round, two reports describe the same structure: an hour of independent work in a recorded session, then an interviewer joins to review it and, in one report, to watch you finish a final part live. Reported problems are a three-part referral network and a two-part task scheduler. In the Applied AI version the first hour allowed Google for syntax but no LLMs, and no tests were provided. One candidate says nobody supervised the candidate room. One 2025 candidate says others around them used AI coding tools freely and advises making sure your Cursor tokens are topped up.
- →Write your own quick tests; one report says none are given.
- →Be able to explain every line in the review. One candidate warns not to let it look like AI did the work.
- →Confirm the AI-tool rules for each round with the coordinator on the day. One candidate's pre-interview email banned LLMs in coding, then the day-of links said they were allowed.
5. Onsite: algorithm round
30 to 45 minutes, discussion only, no code
Candidates talk through an approach and its complexity instead of coding. The most reported problem (four reports) is sorting or finding the top items with a comparison function that is wrong some of the time; follow-ups ask you to minimize the number of calls, then the wall-clock time when each call is slow. Other reports describe finding an exchange rate across currency pairs and searching a friend network using as few API calls as possible.
- →Clarify the error model before proposing anything; reports stress that the comparator's mistakes are repeatable.
6. Onsite: system design
45 minutes
Three reports describe a job scheduler or work orchestration system, and one commenter says knowing the HelloInterview job-scheduler walkthrough was enough to pass it. Other reported designs are a Ticketmaster-style system for one MLE track and a candidate search system for a given job description. One MLE candidate advises matching your depth to the interviewer's background and not explaining domains they do not know.
- →Treat duplicate runs, retries and dependencies as first-class in a job scheduler design.
7. Onsite: candidate search build (SWE and Applied AI)
90 minutes: 75 minutes building with LLMs allowed, then 15 minutes discussion
Two reports describe building a search system that takes a job description and returns the best candidates from a database. In the detailed report you must get it running end to end in 75 minutes, test it against an evaluation endpoint, iterate on the results and submit them; the interviewer then asks how you used LLMs, how you designed it and how you found and fixed problems. That candidate used Claude and GPT heavily with a retrieval and ranking design and suspects their evaluation scores cost them the offer.
- →Use AI tools for speed but have your own design. The report says the expectation is fast AI-assisted coding plus your own judgment.
- →Leave time to iterate on the evaluation results, not just to get it working.
8. Onsite: ML mini-project (MLE)
60 minutes building (a coding agent allowed in one report), then about 30 minutes of review in one report
Two MLE reports describe implementing and training a small autoregressive language model (nanoGPT in one report, trained on Shakespeare and run end to end in the other), followed by transformer fundamentals: architecture, pre- and post-norm, attention, position embeddings and decoding strategies. A third MLE candidate instead had a LeetCode-style problem written in their own IDE projected on a screen, and a commenter in the same thread says their third round was about agentic reinforcement learning.
- →Practise training a tiny GPT end to end against a clock, then review transformer basics for the discussion.
9. Hiring manager or executive chat
Short; 15 minutes in one 2025 report
A September 2026 commenter says someone who passed the onsite went on to meet the hiring manager the same day. In 2025 one candidate had a 15-minute chat with the CTO after passing the technical rounds and was rejected at that step.
- →Bring questions about the company, but read the room; the 2025 candidate felt their pointed business questions landed badly.
How the Mercor loop differs by role
Software engineer
Verbal 30-minute phone screen, then an onsite of coding, a verbal algorithm round, system design and, in one report, the candidate search build. One commenter says a candidate who passed met the hiring manager the same day. A full-stack candidate who passed a similar screen says it focused on reading code for problems and design discussion.
Applied AI engineer
One report has a 20-minute behavioral screen and a three-hour onsite (90-minute referral-network coding and the 90-minute search build); another has a debugging-plus-quick-questions screen; a third started with a four-hour take-home.
Machine learning engineer
One candidate says MLE hiring had two roles, marketplace and frontier data, and a commenter heard from HR that marketplace was folded into frontier data. Two MLE onsites swap the coding round for the language-model mini-project, alongside system design and the comparator algorithm round; a third MLE onsite had a LeetCode-style coding round instead.
Infrastructure, data and research
Single reports each: a 45-minute infrastructure Q&A with about 20 questions (that candidate's thread says they later got an offer), a 20-minute data engineering case study, and a research engineering conversation on post-training data.
After the final round
Results are quick. One MLE candidate heard by email at 9pm after a noon-to-4pm onsite, another candidate heard the same day, and two others were rejected two days after the onsite, one by a template email. One candidate believes they received three no-hire votes, and a commenter says they were sent home after three no-hire rounds. Rejection emails in the reports give no specific reasons. The same candidate who was told the onsite must be in person was asked for three references at that stage.
What does Mercor ask in each round?
Mercor interviews span 4distinct round types, shown below. Counts reflect distinct questions per round across the loops we’ve indexed.
- •On the phone screen: keep every answer short. It is about five questions in 30 minutes, and a candidate who reached only three did not pass. The candidate who passed covered idempotency-key reuse and record-before-charge on payments, and duplicate computation plus lock scope on the cache review
- •On the job-scheduler system design: walk the HelloInterview framework fluently and make distributed correctness first-class — leader election, CAS/idempotency to prevent duplicate runs, and a DAG dependency-counter — rather than presenting a single-node design
- •On the faulty-comparator algorithm round: lead with the key insight that errors are deterministic (so re-calling the same pair never helps), then separately optimize comparison count vs wall-clock time (parallelize independent comparisons when each call is slow)
- •On the 3-part referral network: use cycle-safe traversals, clarify the flow-centrality definition before coding it, and write your own quick tests (the round provides none)
- •Tailor communication to the interviewer — reports say Mercor weights communication fit heavily and advise not going deep on domains the interviewer doesn't share
- •When a round allows a coding agent (some onsite/mini-project rounds do), use it to move fast but be able to explain every line — candidates who pasted AI output they couldn't defend got caught
- •Generating code with an LLM without understanding it — candidates explicitly reported being caught out this way on the coding rounds
- •Designing a single-node job scheduler that ignores leader election, exactly-once/idempotency, or DAG dependency management
- •Going deep on ML-specific content when the interviewer expected general systems knowledge (a reported failing pattern on the SD round)
- •Not clarifying under-specified prompts — the comparator's determinism, the referral flow-centrality definition, and mkdir-style ambiguities are left open on purpose
- •Being thrown by the chaotic logistics (interviewer no-shows, conflicting LLM-usage rules) instead of confirming the ground rules with the coordinator up front
How to prepare for the Mercor interview
1. Drill the phone screen format
Practise answering out loud, in about five minutes each: sorting complexity both ways, a small probability puzzle, reviewing a threaded cache and a CSV writer, and a payment retry scenario. The goal is to finish all of them.
2. Practise the comparator problem
Work through sorting and top-k with a comparison function that is wrong a fixed fraction of the time, then optimize call count and then wall-clock time.
3. Prepare a job scheduler design
It is the most reported onsite design. Cover leader election or locking, duplicate prevention, retries and task dependencies.
4. Rehearse a timed AI-assisted build
Build a small retrieval and ranking search over a dataset in 75 minutes with your AI tools, including an evaluation loop. MLE candidates should instead rehearse training a tiny GPT in an hour.
5. Sort out logistics early
Plan travel to San Francisco, have three references ready, top up your AI coding tool, and confirm the AI rules for each round on the day.
FAQ
How many rounds is the Mercor interview?▾
Is the Mercor phone screen a coding interview?▾
Is the Mercor onsite in person?▾
Can you use AI tools in the Mercor interview?▾
How long does it take to hear back from Mercor?▾
How hard is the Mercor interview?▾
How fresh is this Mercor interview data?▾
Get the full Mercor catalog
Every question, every candidate-reported follow-up, and what passers actually do. Monthly refresh.