What the round looks like
A HackerRank session with your screen shared. The prompt is a simulation: K arms, each paying one with its own probability, T rounds, a random generator seeded from the input, the UCB1 policy (every arm once, then the arm with the best mean plus the exploration bonus), and the total reward printed at the end. You may ask the assistant questions, let its agent mode write the code, and search. The question bank's version of the prompt carries the exact input lines and the formula, and warns that the arm count, the reward setup and the expected output are not fixed from one interview to the next.
Then the questions. The September candidate's summary: the coding is not hard, but there are many follow-ups, and you have to explain the common multi-armed bandit strategies clearly. They did not pass.
Why the simulation is the easy half
An agent writes a bandit loop in a minute. What it cannot do is sit in the chair when the interviewer asks what the square-root term is doing, why it shrinks as an arm is pulled more, what Thompson sampling does instead, or what breaks when the arms' probabilities drift mid-run. Those are the questions the reports describe, and the question bank lists the same set: regret, continuous rewards, a huge number of arms, non-stationary rewards, Thompson sampling to compare.
Our read: the round is an ML conversation with a coding exercise as the ticket in. Use the agent for the loop, then spend your attention on being able to defend every line of it and every alternative to it.
What is confirmed, and what is not
- Confirmed: a final-round coding interview for ML roles on HackerRank with the screen shared; AI assistance allowed; the bandit simulation as the task; the follow-ups on strategies as the bulk of the round. Both reports agree on the shape.
- Reported once: agent mode and web search explicitly allowed; the ML case studies elsewhere in the loop.
- From the question bank, not first-hand: the exact input format, the UCB1 formula, the list of follow-up topics.
- Not known: how the coding and the conversation are weighted, and whether non-ML roles see the same round.
How we would prepare
- Be able to derive UCB1 on a whiteboard. The mean, the bonus, why the bonus has the log of the round and the square root of the pull count, and what happens to it as pulls accumulate.
- Implement Thompson sampling once by hand. A Beta posterior per arm, a sample each, pull the largest. Know what it does better than UCB1 and what it needs that UCB1 does not.
- Have the non-stationary answer ready. A sliding window or a discount on old observations, and the exploration cost of forgetting.
- Make the simulation reproducible on purpose. One generator, one draw per round after the arm is chosen, so the interviewer can check a number. It is the kind of detail the agent gets wrong silently.
Nyrion's Wayfair section starts with Multi-Armed Bandit Simulation, built from these reports: UCB1 reproduced to the last reward on a seeded generator, then Thompson sampling as a second policy on the same contract, then arms that drift and a sliding-window variant that forgets. The task and the follow-up topics are what candidates describe; the function signature in place of standard input, the exact random-draw order and the later stages are ours.
How this round compares with other companies' versions, and what every version grades: the AI coding interview guide.
Frequently asked, briefly answered
Which Wayfair round allows AI?
The final-round coding interview for machine-learning roles, ML scientist and machine-learning engineer, in the two reports we read from June and September 2026. It runs on HackerRank with the screen shared; AI assistance is allowed, including an agent mode that writes the code, and so is searching the web.
What is the task?
A multi-armed bandit simulation: K arms that pay out with given probabilities, T rounds, a seeded random generator, the UCB1 policy, and the total reward as the answer. The question bank's version of the prompt gives the input format and the UCB1 formula and warns that the number of arms, the reward setup and the expected output vary.
Is the coding the hard part?
Not according to the candidates. One wrote that the coding is not hard but there are many follow-up questions, and that you have to explain the common bandit strategies clearly. That candidate was rejected.
What do the follow-ups cover?
The strategies and their trade-offs: what the UCB1 exploration term does, its regret, continuous rewards, a huge number of arms, arms whose probabilities drift, and implementing Thompson sampling to compare. The question bank lists these; the reports confirm the emphasis.
What else is in the loop?
ML case studies heavy on product sense, per the September report. This page covers the coding round only.
Practice problems tagged for Wayfair
Browse the catalog and open the Wayfair section. No login required to look; sign in when you want to solve one.