Google DeepMind Phone Screen: What It Covers and How to Pass

You spent two months on LeetCode. Graphs, DP, the full sliding window situation. You can implement Dijkstra's from memory and explain why merge sort is O(n log n). You feel ready.

Then the DeepMind interviewer asks you to explain how attention works in a transformer. And why it scales quadratically. And what you would do to fix that.

Preparation that covers standard FAANG coding gets you halfway through a DeepMind phone screen. The second half is ML depth, evaluated just as seriously as the algorithms round. Most candidates find out about this gap during the call, not before it.

One Screen or Two?

Most candidates go through one technical phone screen before the onsite. Some tracks, particularly research engineer roles, see two. Each call is 60 minutes.

The platform is CoderPad, not a Google Doc. You can execute code in real time. Interviewers expect you to run it. Typing a solution and staring at it hopefully is not the move.

The call is on Google Meet. No timed take-home in advance. A human interviewer is on the other end, watching you work through problems live. Or more accurately, listening to you work through problems live. Silence is not your friend here.

Your Track Determines What You're Actually Being Tested On

DeepMind hires across three distinct tracks, and the phone screen content shifts meaningfully between them.

Software engineer track: One coding round. Two LeetCode-style problems at medium-to-hard difficulty. Graphs, dynamic programming, hash maps, binary search. This looks the most like a standard Google phone screen, but expect a basic ML fundamentals question appended at the end, even for infrastructure-focused roles. Yes, even if you are building the deployment pipeline.

Research engineer track: Coding plus ML depth. The coding component covers algorithm design at medium-hard level. The ML component tests practical knowledge: implementing a training loop, writing a custom loss function, explaining how attention works and why it scales. You may also be asked to discuss a paper you have actually read recently.

Research scientist track: The phone screen leans heavily on research depth. Coding is present but shorter. Paper discussion is central. You will be asked about methodology, limitations, and how you would extend a specific piece of work. A publication record is expected before you apply.

If you are reading this, you are most likely on the software engineer or research engineer track.

What the Coding Round Actually Tests

The target difficulty is LeetCode medium, with hard problems appearing regularly. Glassdoor reports and candidate accounts from 2024-2025 point to graph traversal, dynamic programming, interval problems, and data structure design as the most common categories.

You have 60 minutes split across everything: clarification, approach discussion, coding, debugging. That gives you roughly 20-25 minutes per problem if two are asked, or a longer single problem with follow-ups.

Common reasons candidates get flagged in written feedback:

Jumping to code before stating the approach out loud
Not running the code when CoderPad lets you run the code
Solving the problem but skipping complexity analysis
Ignoring edge cases until the interviewer raises them

The coding bar at DeepMind is comparable to Google product teams, not lower because of the research angle. Some candidates assume the ML pedigree matters more than the algorithm score. It does not. You need both.

ML Depth Is a Scored Dimension, Not a Bonus Topic

For research engineer roles, interviewers assess machine learning fundamentals in the same call as coding, or in a separate dedicated round.

What gets covered:

Transformers. How attention works, why it computes pairwise dot products across the sequence, what the computational bottleneck is. "It's quadratic in sequence length" is table stakes. Understanding why is the actual test.
Reinforcement learning. DeepMind's work spans AlphaGo, AlphaFold, Gemini. You need to know reward signals, policy gradient, value functions. No published RL research required, but "I've heard of it" is not enough.
Training and debugging. High bias versus high variance. How you would diagnose a model that is not converging. When to change the learning rate versus the batch size. These are bread-and-butter questions that catch people who have only ever touched ML through a Hugging Face pipeline.
Finetuning and RLHF. Supervised finetuning is now table stakes for anyone working near LLMs.
Math. Gradient descent derivation. Matrix multiplication complexity. Why eigendecomposition matters. DeepMind interviews have more mathematical rigor than standard Google product interviews. A lot more.

Candidates who sail through the coding round and stumble on ML depth get rejected. Both halves have to pass.

What the Interviewer Is Actually Scoring

The interviewer submits written feedback after the call. That feedback feeds into a research-heavy hiring committee, which is slower than Google's standard product committee. Expect two to four weeks between phone screen and onsite decision.

The dimensions evaluated in the phone screen:

Algorithm proficiency: correct solution, optimal complexity, clean implementation
ML fundamentals: depth and accuracy of technical reasoning about models and systems
Communication: did you explain your thinking before and during coding, not just after
Reasoning under pressure: clarifying questions, handling follow-ups, recovering from wrong turns

The committee reads the written record, not just the recommendation. Generate quotable evidence during the call. If you said something insightful, the interviewer can write it down. If you were silent for four minutes and then produced a correct solution, there is nothing to write in your favor.

How DeepMind Differs from a Standard Google Screen

The Google technical phone screen follows a well-established format: one coding problem, 45 minutes, algorithm focus. DeepMind's phone screen is similar in structure but broader in scope and evaluated by a committee that includes research staff.

A few concrete differences:

AI tools are prohibited. DeepMind's 2026 policy explicitly bans ChatGPT, Copilot, and any AI assistant during interviews. The irony of an AI company banning AI tools during interviews is intentional. They are measuring what you actually know.

Math rounds exist. Product teams at Google do not run math theory sessions. DeepMind does, especially for research engineer and research scientist tracks.

Paper discussions are standard. Expect to discuss recent research if you are applying to anything above pure SWE. Have one paper you know extremely well, including its weaknesses.

The process is slower. Total timeline from application to decision averages 41 days and can extend to three months. No contact during committee review is normal, not a rejection signal.

Prepare for Both Halves or Prepare to Fail

For coding:

Work through LeetCode mediums across five core patterns: graphs (BFS and DFS), dynamic programming (1D and 2D), hash map and sliding window, binary search on the answer, and heap or priority queue problems. You do not need 300 problems. You need clean, fast execution on 50-75 representative ones with articulate complexity analysis every time.

For ML depth:

Read the original "Attention Is All You Need" paper and be able to explain the architecture piece by piece. Understand gradient descent from first principles. Know the standard training loop in PyTorch: forward pass, compute loss, backward pass, optimizer step. Implement it from memory in 15 minutes.

For RL fundamentals, read the first three chapters of Sutton and Barto. Those concepts show up in DeepMind interviews more than at any other tech company.

For the ML coding section:

Practice implementing a basic attention mechanism, a softmax function without using torch.nn.functional, and a training loop with manual gradient computation. These are common phone screen extensions for research engineer roles.

For communication:

Voice your reasoning before you code, not after. State the approach, its complexity, and why you chose it. Then code. If your prep has been solo LeetCode grinding, you are practicing the wrong half. SpaceComplexity runs AI-powered voice mock interviews that score you on the same dimensions DeepMind's interviewers use, including technical communication and reasoning narration. Getting ten timed reps before the real screen is the difference between candidates who pass and candidates who say "I knew the answer but fumbled it."

Strong Coding Is Not Enough

The most reported rejection reason in post-interview feedback: solid coding, weak ML depth. Candidates preparing for standard FAANG interviews arrive ready for algorithms and underprepared for machine learning. DeepMind's product is AI research. The people interviewing you think about ML systems every day. Shallow answers are easy to detect when you are talking to someone who wrote the paper on it.

The second common mistake is skipping clarification. On the first problem especially, spend two to three minutes defining inputs, outputs, constraints, and edge cases. It is a scored signal, not wasted time.

A third pattern that shows up constantly: candidates who go quiet while thinking. The interviewer is not watching your screen for a solution to appear. They are listening to hear how you reason. Half-thoughts beat silence. Silence gives them nothing to write, and an empty feedback form does not get you to the onsite.

What Happens If You Pass

A passed phone screen triggers scheduling for the virtual onsite loop: four to five rounds across a full day, 45-60 minutes each, covering coding, ML system design, paper discussion (research roles), and behavioral assessment. The full onsite process is covered separately.

Positive interview experience rates on Glassdoor sit around 65%. Interviewers are consistently described as supportive and willing to clarify. The process is hard, but the interviewers are not trying to trick you.

Key Takeaways

One to two technical rounds, 60 minutes each, on CoderPad with executable code
Coding floor is LeetCode medium-hard; graphs, DP, and interval problems appear most
ML depth is a scored dimension for research engineer roles, not a bonus topic
Math and paper discussions are part of the process on research tracks
AI tools are prohibited under DeepMind's 2026 interview policy
Silence reads as confusion; articulate your reasoning throughout
Total process averages 41 days; a slow committee phase is normal
The most common rejection reason is strong coding paired with weak ML fundamentals