Skip to content

ML Job Interviews: The Ultimate Guide

Source: ML Job Interviews: The Ultimate Guide · Silvia Sapora

Overview

Silvia Sapora (PhD in ML, now Research Scientist) wrote the definitive guide to ML research scientist interviews after receiving offers from DeepMind (accepted), Isomorphic Labs, Cohere, Meta, and a startup. The guide covers the entire pipeline: from getting interviews to preparation strategy to technical rounds and behavioral performance.

Getting Interviews

The bar for getting callbacks is high but well-defined:

  • ~3 first-author papers at top ML venues (NeurIPS, ICML, ICLR, AISTATS)
  • Plus a strong internship (or equivalent industry research experience)
  • These two signals are what make recruiters and hiring managers take your application seriously

Sapora's key advice: "If you're already getting interviews, more papers won't help — focus on interview prep instead." Many PhDs make the mistake of trying to publish one more paper to improve their chances, when the marginal return on interview preparation is far higher.

Interview Structure

The ML research scientist interview typically consists of:

  1. Recruiter screen — 15–30 minute chat about background, logistics, timeline
  2. Technical rounds (3–8 rounds) covering:
  3. Coding — LeetCode Medium/Hard, with time pressure
  4. ML coding / debugging — Implement attention, backward passes, flash attention from scratch; debug training pipelines
  5. ML knowledge — Theory, fundamentals, depth in your claimed areas of expertise
  6. System design — Design a training system, deployment pipeline, or data infrastructure at scale
  7. Behavioral rounds — Often underestimated; these matter more than candidates think

Interview Preparation

Time Allocation

Sapora recommends at least one month of dedicated preparation. This is not something to cram in a weekend.

Mock Interviews

The most practical recommendation: paste the full role description into an LLM (she recommends Claude) and run a mock interview. The LLM can ask technical questions, probe your reasoning, and give feedback. This is substantially more effective than studying alone because it simulates the actual interview dynamic.

LeetCode Strategy

  • Blind 75 — Start here for fundamental patterns
  • NeetCode 150 — Work through for breadth
  • Focus on Mediums — Aim to solve them in 20 minutes or less
  • Hards matter less for ML roles than for SWE roles, but be prepared

Books to Study

  • "Designing ML Systems" by Chip Huyen — For ML system design rounds
  • JAX Scaling Book — For understanding modern training infrastructure

Courses

  • Gilbert Strang's Linear Algebra — Watch at 2x speed for a quick refresher on the math that matters

ML Coding Baseline

You should be able to implement from scratch without looking up references:

  • Attention — Scaled dot-product, multi-head, causal masking
  • Backward passes — Manual backprop for common operations
  • Flash attention — The tiling algorithm, IO-aware implementation

Don't Underestimate Behavioral Rounds

Sapora's strongest caveat: behavioral rounds can make or break your offer. Technical competence is table stakes — behavioral rounds test whether you are someone the team wants to work with. She advises preparing concrete stories for:

  • Conflict resolution with colleagues
  • Failed projects and what you learned
  • Times you took initiative beyond your role
  • How you handle vague/ambiguous problems
  • Your research philosophy and motivations

At top labs, behavioral rounds often carry more weight than candidates realize because every interviewee is technically strong — the differentiator is whether you seem like a good collaborator.

Key Takeaways

  1. Getting interviews requires ~3 first-author papers at top ML venues plus a strong internship.
  2. If you're already getting callbacks, stop trying to publish more — focus on interview prep.
  3. Allocate at least one month of dedicated prep time.
  4. Use LLMs for realistic mock interviews — paste the job description and let the LLM interrogate you.
  5. LeetCode: Blind 75 → NeetCode 150, focus on Mediums in ≤20 minutes.
  6. ML coding baseline: implement attention, backward passes, and flash attention from scratch.
  7. Study "Designing ML Systems" (Chip Huyen) and the JAX Scaling Book.
  8. Don't underestimate behavioral rounds — at top labs they often determine who gets the offer.
  9. Behavioral stories should cover: conflict, failure, initiative, ambiguity, and motivation.

References