Company interview guide
Cohere Machine Learning Interview Guide (2026)
Enterprise LLM lab: post-training and retrieval design, numpy ML coding, a research talk and, for some teams, AI-assisted coding.
Cohere's machine learning interview loop has 9 reported stages, from recruiter call to behavioral & leadership.
Roles covered: ML Engineer, Research Scientist, LLM Engineer and Forward Deployed Engineer. The question bank holds 21 questions reported from Cohere interviews, 5 of them free to solve in your browser.
- Stages
- 9
- Reports linked to bank questions
- 25
- Bank report dates
- Feb 2023–Sep 2026
- Questions in bank
- 21 (5 free)
- Last reviewed
- October 2026
How the Cohere interview loop works
Cohere's hiring page lists an application review, a recruiter conversation, an optional take-home, a hiring manager meeting and a final round with a few team members. In recent reports the hiring manager conversation usually comes first or right after the recruiter, a take-home or online assessment appears in some loops, and the final round is three or four technical interviews: coding (often with an AI assistant allowed), LLM or ML system design, ML breadth and depth, and a research talk or leadership conversation. Forward-deployed loops swap the research talk for a debugging round. Several candidates describe disjointed scheduling, multiple recruiters and long silences after the final round, so follow up politely and keep other processes moving.
The Cohere interview process, stage by stage
01Recruiter call
Short call
A conversation about your background, the role and logistics. Cohere hires into specific teams: one candidate was referred to a better-fitting role, and another applied to an audio team but was interviewed on LLM topics, so be clear about the area you want.
What it tests
- Why Cohere and why enterprise AI
- Which team and model area fits your experience
- Logistics: location, level, timeline
How to prepare
- Ask which rounds you will get, whether there is a take-home and whether AI coding tools are allowed; the answers differ by team.
- Know Cohere's main product lines (generation, embeddings, reranking) well enough to say where you would contribute.
Recruiter or HR screen first in 5 ML, MTS and FDE reports since 2022 plus one non-technical report; on the company's hiring page. Several 2025-2026 MTS reports start directly with the hiring manager.
02Hiring manager screen
Video call with the hiring manager or team lead
The first real gate in most recent reports, and for several MTS candidates the first call of all. Expect a deep dive into one or two past projects and questions about how your work maps onto the team, including customer-facing experience for applied roles.
What it tests
- Depth and ownership behind the projects on your resume
- Fit with the team's model area (for example efficiency, post-training, retrieval, audio)
- Experience working with customers or product teams
- Clear communication of technical trade-offs
How to prepare
- Prepare a five-minute walkthrough of your strongest project, with the decisions you made and the numbers that moved.
- Expect broader questions about your current employer's strategy; answer briefly and steer back to your own work.
Hiring manager conversation, usually a project deep dive, in 8 reports (MTS, MLE, FDE, one intern), mostly 2024-2026. The hiring page lists it after the take-home; reports put it before.
03Take-home or online assessment (some loops)
A take-home project (one report says AI coding tools were expected) or, in one forward-deployed loop, an online coding assessment; format varies by team
Not every loop has one. Candidates report a coding take-home before the final round for a member-of-technical-staff role and a research scientist role, and an online assessment for a forward-deployed role. A public write-up describes a take-home to build a small LLM agent over business data that must refuse personal-data requests.
What it tests
- Turning a loose brief into a working, testable deliverable
- Guardrails: refusing unsafe requests and limiting what the model can see
- Evaluation design: accuracy, safety and reasoning as separate scores
- Code you can explain and defend in a follow-up
How to prepare
- Ask whether AI tools are allowed and what the follow-up discussion covers before you start.
- If AI tools are allowed, keep the code small enough that you can defend every line in the follow-up.
Coding take-home in 2 reports from 2025 (MTS, research scientist), an online assessment in a 2026 FDE report, a coding exam in a 2022 intern report. Hiring page calls it optional; the agent brief is from a public write-up.
04Coding: ML coding and AI-assisted coding
Live, in a notebook or shared editor; one intern task ran 45 min; some loops allow an AI assistant, a software-engineering loop banned it
The ML coding round asks for small LLM components in plain numpy, checked by asserts against fixed outputs, followed by concept questions. A newer format gives you a few hundred lines of an existing program and asks you to rewire it around LLM calls with an AI assistant available, while judging whether you understand what you ship. Intern loops used a timed notebook task to write a benchmark evaluation for a language model.
What it tests
- Greedy, top-k and top-p sampling written by hand, matching expected token sequences
- Attention and tensor manipulation without a framework
- Reading an unfamiliar codebase and finding the seam to change
- Treating model output as untrusted input: parsing, validation, fallbacks
- Writing a small evaluation harness for a language model benchmark
How to prepare
- Confirm shapes, tie-breaking and when the random generator is called before coding; fixed-sequence asserts punish small differences.
- With an AI assistant, read and explain each suggestion before accepting it; accepting code you cannot explain has been called out as a negative signal.
- Practise numpy decoding and a causal attention forward pass until you can write them in under 15 minutes.
Free reported questions for this stage
- Greedy, Top-k and Top-p Decoding in NumPyML CodingMediumSolve
- Swap a Text Game's Heuristics for LLM CallsLLMs & GenAIMediumSolve
- Multi-Head Causal Self-Attention From ScratchML CodingHardSolve
Coding round in 9 reports: AI-assisted coding in 3 from 2025-2026, numpy decoding in a 2026 onsite, timed notebook eval tasks in 2 intern reports, attention or tensor code in 2 older ones.
05LLM / ML system design
Open discussion with an interviewer
The most consistently reported technical round in recent loops. Prompts ask you to build a model capability end to end: collect data, run supervised fine-tuning and reinforcement learning, evaluate and serve, often for a single enterprise customer. Forward-deployed loops ask for AI product designs such as an LLM code review system.
What it tests
- Post-training pipelines: data collection, SFT, reward models, RL, evaluation
- Why an RL or preference stage is needed at all, and how rewards are chosen
- Model behavior design, such as knowing what lies past its training cutoff
- Evaluating RAG and retrieval systems without labelled data
- Enterprise constraints: private data, deployment, feedback loops
How to prepare
- Have one end-to-end post-training blueprint ready and adapt it to coding, math or behavior prompts.
- Expect the interviewer to push hard on how each dataset is built; be specific about sources, filtering and labels.
- Say how you would measure success before going deep on training details.
Free reported questions for this stage
- Post-Training a 7B Model for One Enterprise's CodeLLMs & GenAIHardSolve
- Evaluating RAG Systems, RAG Components and the TransformerLLMs & GenAIMediumSolve
ML or LLM system design round in 6 reports from 2025-2026 (MTS, MLE, FDE): post-training pipelines, a math-solving model, an LLM code review system. A software-engineering report also had one.
06ML breadth & depth
Conceptual questions with follow-ups
A fundamentals round that goes from transformers and statistics into model architecture. Retrieval and embedding questions framed as an enterprise customer's problem with numbers attached appear in a compiled public list, so expect quick cost and capacity arithmetic as well.
What it tests
- How transformers work, and statistics fundamentals
- Bi-encoders versus cross-encoders and where a reranker pays off
- Mixture-of-experts parameter versus compute trade-offs
- Embedding compression: dimensions, quantization and recall
- When to prompt, retrieve or fine-tune
How to prepare
- Be ready for deep questions on statistics and transformer internals, not only LLM buzzwords.
- Practise computing index memory, reranker throughput and active parameter counts on the spot; retrieval is central to Cohere's products.
ML breadth or fundamentals round in 2 reports from 2025 (MTS, research scientist) and 3 from 2022-2023 (MLE, intern). Retrieval and embedding examples come from a compiled public list.
07Problem solving: incident debugging (forward-deployed track)
Interactive: you are given a system design and a user-facing problem and investigate with the interviewer
Forward-deployed loops include a simulated incident in which you narrow down a root cause by asking for logs and metrics; a software-engineering candidate describes the same format. One candidate was rejected on this round alone despite feeling it went well, so treat it as a full round, not a formality.
What it tests
- Forming and testing hypotheses in a sensible order
- Asking for the right signal: metrics, logs, recent changes
- Diagnosing retrieval quality regressions and embedding drift
- Communicating the root cause and a fix
How to prepare
- Narrate a plan first (what changed, where the symptom appears, what you would check), then drill down.
- Rehearse RAG regressions: index growth, duplicate or stale documents, ANN settings, embedding model changes.
Debugging or problem-solving round in both forward-deployed reports (2026); a 2026 software-engineering report describes the same incident simulation. Not reported for research or MTS loops.
08Research talk
Presentation of your past work or a published paper, then questions
Research and member-of-technical-staff loops include a talk on your own work; one research scientist loop had both a research discussion and a research deep dive. One report says the interviewer may already know the paper, skip the slides and go straight to weaknesses in the setup, real-world usefulness and what the field has learned since.
What it tests
- Whether your experimental setup holds up
- Whether the results are practical outside the paper's setting
- Awareness of later work that changes your conclusions
- Your personal contribution on joint work
How to prepare
- Critique your own paper before the interview and prepare an honest answer for each weak point.
- Read the follow-up work since publication and say how it changes your baselines or claims.
Research talk, paper presentation or research deep dive in 4 reports from 2025-2026 (2 MTS, 1 MLE, 1 research scientist), consistent on a hard Q&A about your own work.
09Behavioral & leadership
Conversation, sometimes labelled leadership
A conversation about how you work: ownership, speed, collaboration and dealing with customers. Cohere's stated values are momentum, openness and autonomy.
What it tests
- Ownership and moving quickly with incomplete information
- Working across research, product and customer teams
- Leading or influencing without formal authority
- Handling disagreement openly
How to prepare
- Prepare stories that show you shipping under ambiguity and owning the outcome.
- For applied roles, have one concrete example of working directly with a customer.
Behavioral or leadership round named in 3 reports (2025 MTS, 2025 research scientist, 2022 intern). The company's hiring page states values but not a separate behavioral round.
Questions and themes reported most often
- Present and defend your own research (4 reports, research talk)
- Hiring manager deep dive on past projects (4 reports, hiring manager screen)
- Coding with an AI assistant allowed (3 reports, coding: ml coding and ai-assisted coding)
- Incident or system debugging with the interviewer (3 reports, problem solving: incident debugging (forward-deployed track))
- Design a post-training pipeline (data, SFT, RL, eval) for a target skill (2 reports, llm / ml system design)
- Code transformer attention or other tensor manipulation (2 reports, coding: ml coding and ai-assisted coding)
- Timed notebook task: write an evaluation for a language model (2 reports, coding: ml coding and ai-assisted coding)
- Implement LLM decoding (greedy, top-k, top-p) without a framework (1 reports, coding: ml coding and ai-assisted coding)
What Cohere asks about
- LLM / ML system design28%
- ML coding22%
- ML & LLM concepts20%
- Research talk & behavioral12%
- General & AI-assisted coding12%
- Debugging & problem solving6%
Cohere interview prep checklist
- Can design a post-training pipeline for a 7B model and justify the RL stage
- Wrote greedy, top-k and top-p decoding in numpy that matches fixed expected outputs
- Implemented multi-head causal attention from scratch
- Can evaluate a RAG system for a customer with no labelled data
- Rehearsed a hostile Q&A on your own paper or main project
- Practised rewiring an unfamiliar codebase around LLM calls with an assistant, explaining every change
- Have two stories on ownership and customer-facing work
Tips
- Build one end-to-end post-training design (data, SFT, reward model, RL, evaluation, serving, feedback loop) and rehearse it for coding, math and behavior prompts.
- Write sampling loops and causal attention in plain numpy until they are quick and exactly reproducible.
- Prepare your research talk as a defense, not a summary: list the weak points of your setup and the later work that challenges it.
- Get fluent in retrieval: bi-encoders, rerankers, hybrid search, embedding compression and evaluation without labels.
- Ask the recruiter for every round's format and the AI tool policy; it is allowed in some loops and banned in others.
- When using an AI assistant, explain and test what it writes; understanding is what is graded.
- For forward-deployed and software roles, practise talking through an incident from symptom to root cause.
- Expect slow or uneven scheduling and follow up in writing after each stage.
See the rest of the Cohere loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the Cohere loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Cohere interview FAQ
How many interview rounds does Cohere have for ML roles?
Candidates report 9 stages: Recruiter call, Hiring manager screen, Take-home or online assessment (some loops), Coding: ML coding and AI-assisted coding, LLM / ML system design, ML breadth & depth, Problem solving: incident debugging (forward-deployed track), Research talk and Behavioral & leadership. Cohere's hiring page lists an application review, a recruiter conversation, an optional take-home, a hiring manager meeting and a final round with a few team members. In recent reports the hiring manager conversation usually comes first or right after the recruiter, a take-home or online assessment appears in some loops, and the final round is three or four technical interviews: coding (often with an AI assistant allowed), LLM or ML system design, ML breadth and depth, and a research talk or leadership conversation. Forward-deployed loops swap the research talk for a debugging round. Several candidates describe disjointed scheduling, multiple recruiters and long silences after the final round, so follow up politely and keep other processes moving.
What does Cohere ask in ML interviews?
By reported question mix: LLM / ML system design 28%; ML coding 22%; ML & LLM concepts 20%; Research talk & behavioral 12%; General & AI-assisted coding 12%; Debugging & problem solving 6%. The most repeated themes: Present and defend your own research; Hiring manager deep dive on past projects; Coding with an AI assistant allowed.
How should I prepare for the Cohere ML interview?
Build one end-to-end post-training design (data, SFT, reward model, RL, evaluation, serving, feedback loop) and rehearse it for coding, math and behavior prompts. Write sampling loops and causal attention in plain numpy until they are quick and exactly reproducible. Prepare your research talk as a defense, not a summary: list the weak points of your setup and the later work that challenges it.
Are there free Cohere interview questions?
Yes. LeetML has 21 questions reported from Cohere interviews, and 5 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- AnthropicA timed, level-by-level build exercise, practical live coding, LLM infrastructure design and a values interview that really counts.
- OpenAIPractical coding, hands-on ML debugging and LLM system design, with a strong mission check.
- Mistral AIAttention coded from scratch in PyTorch, a heavy LLM fundamentals quiz and practical work against the Mistral API.
- Together AIInference performance depth, serving and scheduler design, and practical systems coding for an open-model AI cloud.
- Google DeepMindAlgorithmic coding and broad maths, stats and ML fundamentals, then team, research and culture conversations.
- Hugging FaceRemote-first open-source ML company: a motivation-heavy first call, a take-home for some roles, then conversational technical rounds.