Jobra
🐞Step 1 of Jobra:AI Debugging Rounds
DSA+AIPreparing you for both.
JOBRA · AI DEBUG ROUND (PREVIEW)sum.cpp
// buggy snippet
int total(vector<int>& a, int n) {
    int sum = 0;
    for (int i = 0; i <= n; i++) { // 🐞        sum += a[i];
    }
    return sum;
}
A real codebase. A real bug. You + AI.

Jobra drops you into a broken repo. Find it, fix it, ship it. AI helps you debug like on the job.

Prompt:Correct the loop condition
Tap 'Let Jobra fix it' below to simulate AI debugging...

Meta Machine Learning Engineer Interview Guide: Coding, AI Coding and ML System Design

2 min read

Prepare for Meta machine learning engineer interviews with reported assessment patterns, interview questions, and preparation advice.

In this article
On this page

Last reviewed November 15, 2026.

For machine learning engineer candidates.

Built by the Codinzhub team from Meta MLE interview experiences that candidates shared. Each experience is one person's interview, and only a few were detailed, so use this as a map, not a promise.

The 30-second version. Screens on core ML knowledge (attention and its variants) and coding, then an onsite of about four rounds: coding, AI-enabled coding, ML system design and behavioral. Teams in GenAI also asked research-specific rounds. Some interviewers asked hard variants despite recruiters saying they would not.

Round-by-round breakdown

Stage 1: Application, where candidates applied

How it works. Candidates applied online to search, ads and recommendation roles, or were matched to GenAI teams. One process was disrupted by a reorganisation mid-loop.

Stage 2: Screens

At a glance

  • Length: about 45 minutes each
  • Questions per round: ML concept questions or 1 to 2 coding problems
  • What was asked: attention variants, a robot cleaner variant
  • Most common: no single question repeats

Questions candidates reported:

Question (our summary)Follow-up reported
Robot that must find a target in an unknown room through a move and turn APIRestore position and direction when backtracking
Core ML: attention and its variantsDepth on each

Stage 3: The onsite

Coding and AI-enabled coding

At a glance

  • Length: 45 to 60 minutes
  • Questions per round: 1 or 2 problems
  • What was asked: permutations, word subsets
  • Most common: the maximum unique-letter word subset

Questions candidates reported:

Question (our summary)What it tests
The second-largest number you can form from a list of digitsPermutations
Largest set of words with no shared letters (AI-enabled)Bitmasks, pruning

More questions: Browse Meta questions.

ML system design

At a glance

  • Length: about 45 minutes
  • Questions per round: 1 design
  • What was asked: recommendation and ranking systems, assistants with safety measures
  • Most common: nearby place recommendation

Questions candidates reported: nearby place recommendation for users on the move; notification ranking; a conversational assistant with safety measures; MoE pre-training and async RL post-training for GenAI teams.

Behavioral

Questions candidates reported: your proudest project, then challenges and pushback on it; leadership and culture questions with directors for GenAI roles.

More questions: Browse Meta questions.

The Codinzhub Next-Round Playbook

What MLE candidates who moved on did. Put together by the Codinzhub team from the MLE experiences in this guide.

  • Revise attention, transformers and their variants for the screens.
  • Practise ML system design with safety, ranking and serving trade-offs.
  • In the AI-enabled round, do not trust the AI for the optimal answer; one candidate said neither model they tried found it.

About this guide

This guide is researched and written by the Codinzhub team. We read every Meta MLE interview experience we could find, kept only what candidates actually described, and rewrote it in our own words. Found something out of date? Share your own experience on Codinzhub and we will update the guide.

Ready for the next topic?

Keep track of what you’ve learned.