Company interview guide

ByteDance Machine Learning Interview Guide (2026)

Sequential one-hour rounds that pair a deep resume grilling with a classic algorithm problem and ML fundamentals.

ByteDance's machine learning interview loop has 8 reported stages, from recruiter or hr call to hiring manager and hr rounds; typical length: 2-8 weeks.

Roles covered: ML Engineer, Research Scientist, Applied Scientist and LLM Engineer. The question bank holds 23 questions reported from ByteDance interviews, 7 of them free to solve in your browser.

Try a free question

Interview process length
2-8 weeks
Stages
8
Reports linked to bank questions
23
Bank report dates
Dec 2020–Oct 2026
Questions in bank
23 (7 free)
Last reviewed
October 2026

How the ByteDance interview loop works

A short recruiter or HR call, an online assessment for some intern, new-grad and data candidates, then usually two to four technical rounds of 45-60 minutes, each gated on the last and often a week or two apart; some research loops ran six or seven rounds. Research tracks may add a job talk of 60-90 minutes. A hiring manager or director round and an HR round often close the loop. Timelines are uneven: some candidates finished in two weeks, others report long gaps between rounds or cancelled headcount.

The ByteDance interview process, stage by stage

01Recruiter or HR call

Short phone or video call, about 20-30 min

A brief call on your background, motivation, location and work authorization, often with an outline of the team and the rest of the process. Many candidates are approached by recruiters directly; some data-track loops opened with a hiring manager screen instead.

What it tests

  • Why ByteDance and why this team
  • How your past work matches the team's product area
  • Logistics: location, level, timeline, interview language

How to prepare

  • Ask which team you are interviewing for and what it builds; questions in later rounds follow the team's product.
  • Ask whether rounds will be in English or Mandarin, and request English if you need it.

An initial recruiter or HR call in 8 ML-role and 4 data-role candidate reports (shared by all tracks); the company's apply page says you meet your recruiter first.

02Online assessment (mainly new grad and intern)

60-120 min proctored online test, 3-5 problems, sometimes with SQL or multiple choice

Reported mostly for intern, new-grad and data-science candidates, and mostly in 2020-2021 reports; experienced ML hires usually go straight to technical rounds. Problems are standard algorithm tasks from easy to hard.

What it tests

  • Speed on easy and medium array, string, tree and DP problems
  • Getting a hard problem to pass within the time limit
  • Basic SQL or systems multiple choice in some variants

How to prepare

  • Practise timed sets of three or four problems in one sitting.
  • Clear every easy problem first; one candidate says passing needed all of them solved.

Timed online test in 6 ML-role and 5 data-role reports, only 2 since 2024; the company's early-careers FAQ says technical roles usually take an online coding assessment, and its main FAQ mentions tests.

03Technical rounds: live coding

20-30 min of most 45-60 min technical rounds, before or after the project discussion; one problem, sometimes two

Nearly every technical round includes a classic algorithm problem, usually medium, sometimes hard. Interviewers expect the optimal approach in time and space, sometimes without running the code, and may push for a bound tighter than the obvious one. Research loops ask it too, usually at an easier level.

What it tests

  • Prefix sums with hashing and modular counting
  • Binary search on answers and on partitions
  • Heaps, sweep lines and greedy scheduling
  • Dynamic programming on arrays, including sign-flipping products
  • Grid and tree BFS or DFS

How to prepare

  • State the optimal complexity before you code; a correct but slower answer is often not enough.
  • Practise writing clean code in a plain editor without running it, then trace one example by hand.
  • Project questions often overrun, so be ready to solve a medium problem in 15-20 minutes.
  • Prepare for follow-ups that tighten the bound, such as beating O(K log N) by counting levels instead of simulating.

Free reported questions for this stage

A live algorithm problem in 54 ML, research and applied candidate reports, 27 since 2024; research rounds use easier problems. An official campus page cites 3-5 technical rounds (more than most reports) but not their topics.

04Technical rounds: project deep dive and ML fundamentals

Most of each technical round: 20-40 min on your resume, with spoken ML questions drawn from it

Each technical round spends a long stretch on your own projects, often re-asking about the same project in more detail, and turns to fundamentals prompted by the models you mention. Classic ML theory still appears, and transformer and LLM questions have grown more common since 2024.

What it tests

  • Depth on every model, loss, metric and method named on your resume
  • Overfitting, bias-variance and why boosted trees generalize better than one deep tree
  • Logistic regression loss, gradient and convexity; SVMs and AUC
  • Transformer blocks, attention cost, long sequences and FlashAttention
  • Why LLMs use LayerNorm or RMSNorm, and how multimodal LLMs are built

How to prepare

  • For each project, be ready to explain the model, the loss, the metric and why you chose them.
  • Review XGBoost's regularization knobs and be able to say when it still overfits.
  • Know one model family's technical reports well enough to say what each generation changed.

Free reported questions for this stage

Resume deep dive or ML concept questions in 55 ML-role candidate reports, 27 since 2024; 26 of them name specific ML or LLM concept questions.

05Technical rounds: ML coding

Replaces the algorithm problem in some rounds; Python with numpy or PyTorch, sometimes on a whiteboard, about 30 min for a full decoder

Some teams ask you to write core ML pieces from scratch instead of, or after, an algorithm problem. Recent prompts include a full decoder-only transformer in about 30 minutes with no starter code, attention written on a whiteboard, and an MLP backward pass in numpy for research roles.

What it tests

  • Multi-head causal self-attention with correct shapes and masking
  • A small decoder-only transformer forward pass end to end
  • Embedding similarity with edge cases such as empty text and unknown words
  • Exact nearest-vector search and an approximate index for large N

How to prepare

  • Be able to write attention and a decoder block from memory, narrating tensor shapes as you go.
  • Practise in plain numpy as well as PyTorch, since some interviewers do not allow libraries.
  • Research tracks have asked for an MLP forward and backward pass in numpy; drill it.

Free reported questions for this stage

ML implementation tasks in 7 candidate reports, 5 since 2024: attention or a decoder (2), MLP forward and backward (2), embedding similarity, vector search and numpy or PyTorch tasks.

06ML system design

30-45 min inside a technical round, often the second; sometimes alongside a coding problem

Prompts track the team's product: ads re-ranking, abuse and crawler detection, LLM agents for enterprise users, or large-scale model training. Interviewers have steered candidates toward specific techniques, for example graph features for marketplace fraud.

What it tests

  • Framing a detection problem: labels, candidate generation, features and precision versus recall
  • Graph, temporal and account-linking signals for fraud and abuse
  • Recommendation and ads ranking: retrieval, ranking, re-ranking and online metrics
  • Agent design: tool failures, long-running state and offline and online evaluation

How to prepare

  • For abuse problems, reach for a graph of users, sellers and items early; a pure recommender framing has failed.
  • Prepare one recommendation or ads ranking blueprint in depth; it matches many ByteDance teams.
  • Name the offline metric, the online metric and how you would get labels before going deep.

Free reported questions for this stage

ML system design in 6 ML-role candidate reports, all since 2025, each tied to the team's product; data-science loops add recommendation and content-safety cases.

07Research discussion and job talk

Research-scientist tracks: paper discussions inside technical rounds, sometimes plus a 60-90 min presentation

Research loops spend much of each round on your papers and on where your field is heading, and some add a full job talk. One candidate had a round billed as coding turn into a 90-minute talk; another loop had a maths and statistics screen.

What it tests

  • Explaining your research clearly to someone outside your subfield
  • Defending experimental design and baselines
  • Views on the state of the art and open problems
  • Brainstorming research ideas relevant to the team

How to prepare

  • Prepare a deck longer than an hour with a research map, so you can steer toward your strongest work.
  • Have short and long versions of each paper ready, and still prepare for a coding problem.

Paper or research discussion in 14 research-track candidate reports, 5 since 2024; a separate job talk appears in only 3 of them.

08Hiring manager and HR rounds

A manager or director round near the end, then an HR conversation

The manager round revisits your resume and introduces the team's projects; in research loops a director may weigh your background heavily. The HR round covers culture fit, expectations and the offer. Behavioral questions are light elsewhere in the loop.

What it tests

  • Fit between your experience and the team's current projects
  • Motivation, career goals and expectations
  • Questions about the team and role

How to prepare

  • Read about the team's product and be ready to say how you would contribute in the first months.
  • Use the manager round to ask what the team ships and how the role is scoped; several managers spent the round introducing the team.

A hiring manager, director or final HR round in 10 ML-role candidate reports, 4 since 2024; data-role loops show the same HR close. The company's large-model campus page also lists one HR interview.

Questions and themes reported most often

  • Detailed deep dive on your own projects, repeated in every round (47 reports, technical rounds: project deep dive and ml fundamentals)
  • Research talk or detailed discussion of your papers (14 reports, research discussion and job talk)
  • Transformer and LLM concepts: architecture, long context, FlashAttention, normalization (7 reports, technical rounds: project deep dive and ml fundamentals)
  • Overfitting, bias-variance and boosting versus one deep tree (5 reports, technical rounds: project deep dive and ml fundamentals)
  • Abuse detection design: self-dealing sellers, crawlers, unsafe content (3 reports, ml system design)
  • Classic ML theory: logistic regression convexity, SVMs, AUC (3 reports, technical rounds: project deep dive and ml fundamentals)
  • Linked list problems: reversal, k-group variants, adding numbers (3 reports, technical rounds: live coding)
  • Recommendation or ads re-ranking system design (2 reports, ml system design)

What ByteDance asks about

  • General coding34%
  • Resume & project deep dive24%
  • ML & LLM concepts16%
  • ML coding8%
  • ML system design8%
  • Research discussion6%
  • Behavioral & HR4%

ByteDance interview prep checklist

  • Can walk through two projects for 30 minutes each and defend every modelling choice
  • Solved median of two sorted arrays and a prefix-sum counting problem at optimal complexity without running code
  • Wrote multi-head causal attention and a decoder-only forward pass from scratch in about 30 minutes
  • Can explain why boosted trees overfit less than one deep tree, and when they still overfit
  • Can explain LayerNorm versus RMSNorm and how image inputs enter a multimodal LLM
  • Have a ranking or abuse-detection design ready with labels, features and metrics
  • Know which team you are joining, what it ships and why you want it

Tips

  • Treat the resume deep dive as the core of every round: rehearse each project's model, data, metric, trade-offs and your own contribution.
  • Drill classic algorithm problems to the optimal solution; interviewers rarely accept a working but slower answer.
  • Expect each round to be judged on its own and to gate the next, so do not coast after a strong first round.
  • Be able to write attention and a small decoder from scratch in about 30 minutes while explaining shapes.
  • Refresh classic ML theory alongside LLM topics; boosting, regularization, logistic regression and AUC still come up often.
  • Prepare a system design in the team's product area, such as recommendation, ads, e-commerce risk or LLM agents.
  • Confirm the interview language and time zone early; some rounds are in Mandarin and US slots can fall late at night.
  • Keep answers tight, since late starts and long project questions often leave little time for the coding problem.
  • Research candidates should prepare a paper walkthrough for every round and a talk deck in case a round becomes a job talk.

See the rest of the ByteDance loop

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

Unlock with Pro

Practice the ByteDance loop

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

ByteDance interview FAQ

How many interview rounds does ByteDance have for ML roles?

Candidates report 8 stages: Recruiter or HR call, Online assessment (mainly new grad and intern), Technical rounds: live coding, Technical rounds: project deep dive and ML fundamentals, Technical rounds: ML coding, ML system design, Research discussion and job talk and Hiring manager and HR rounds. A short recruiter or HR call, an online assessment for some intern, new-grad and data candidates, then usually two to four technical rounds of 45-60 minutes, each gated on the last and often a week or two apart; some research loops ran six or seven rounds. Research tracks may add a job talk of 60-90 minutes. A hiring manager or director round and an HR round often close the loop. Timelines are uneven: some candidates finished in two weeks, others report long gaps between rounds or cancelled headcount.

How long does the ByteDance ML interview process take?

Reported timelines run 2-8 weeks.

What does ByteDance ask in ML interviews?

By reported question mix: General coding 34%; Resume & project deep dive 24%; ML & LLM concepts 16%; ML coding 8%; ML system design 8%; Research discussion 6%. The most repeated themes: Detailed deep dive on your own projects, repeated in every round; Research talk or detailed discussion of your papers; Transformer and LLM concepts: architecture, long context, FlashAttention, normalization.

How should I prepare for the ByteDance ML interview?

Treat the resume deep dive as the core of every round: rehearse each project's model, data, metric, trade-offs and your own contribution. Drill classic algorithm problems to the optimal solution; interviewers rarely accept a working but slower answer. Expect each round to be judged on its own and to gate the next, so do not coast after a strong first round.

Are there free ByteDance interview questions?

Yes. LeetML has 23 questions reported from ByteDance interviews, and 7 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources