TikTok MLE Interview Experience (2026) - Recommendation Deep Dive & Hard DP, Offer with Sponsorship

• Describe the candidate generation model. What architecture did you use, what were its parameters, what was training cost, and what was inference latency
• What metric did you use in offline evaluation, and how did that correlate to the online A/B metric
• How was the ranker trained and what feature groups did it consume
• Scenario: a new user joins with no history and the candidate generator has nothing to retrieve. What do you do
• Scenario: a new item launches and the ranker has no interaction data. What do you do

For the new-user cold start I went with demographic-bucket fallback rankings plus bandit exploration for the first 20 impressions. For new-item cold start I talked about content-based embeddings from the item metadata as a bridge until click data accumulates. The HM pushed on the bandit part specifically: "what is your exploration budget, and how do you stop the exploration bleeding into the recommendation quality?" I talked about Thompson sampling with a decaying prior and epsilon schedules keyed off per-item impression counts.

The last 15 minutes were a coding problem at the medium level, directly tied to recommendation systems. The shape: given a stream of user events (each a (user id, item id, event type) tuple), maintain the top K most-clicked items per user with efficient updates. I used a per-user min-heap of size K keyed by click count, keyed by a hash table of (user id, item id) - count . When a click arrives, increment the count; if the item is already in the user's heap, resift; if not and the heap is full, compare against the heap top. Standard top-K-with-updates pattern.

[[problem/317?company=22|Recommendation System Model Selection]]

Round 3: Department Head (75 min)

This round was the filter. The Head opened by saying he was going to be pickier on details and that was not a warning, it was a fact. Three parts:

1. Model mathematics deep dive (25 min)
2. Full recommendation pipeline walkthrough (20 min)
3. Hard coding problem (15 min plus 15 min discussion)

Math Deep Dive

He pulled up my chatbot project and asked me to derive, on the fly, the loss function I had used (a contrastive objective), the gradient of that loss with respect to the output logits, and the reason the gradient form avoided a particular instability. This was the most demanding 25 minutes of any interview I have done. He did not want an intuition; he wanted the chain rule explicit. I had reviewed this material the night before and it was still tight.

Takeaway: if you list a model on your resume, know the loss, the gradient form, and at least one numerical stability consideration. Anything less will be noticed.

[[problem/315?company=22|Transformer Architecture Deep Dive]]

Pipeline Walkthrough

He asked me to walk end to end from user request to served recommendation, including which service owns which piece. I drew candidate generation, pre-rank filter, heavy rerank, business rules filter (diversity, freshness, brand safety), and impression logging. The follow-up question was on feature lookup latency: "you said rerank features are fetched from a feature store; what is the p99 of that fetch and what do you do if it exceeds your SLO?" Real answer was a feature cache colocated with the rerank service, with a stale-while-revalidate pattern. He seemed satisfied.

Hard Coding Problem

A dynamic programming problem on arrays with a constraint I had seen in a similar form on LeetCode. I will not spoil the exact problem because it was obviously from a question bank, but it was in the "partition the array into subsegments to maximize a score" family, and the state was dp[i][j] where j encoded a segment count. Time limit was 15 minutes. I had done enough of these that I could derive the state space in 2 minutes and had working code at minute 9. The Head explicitly said "nice" when the code ran on his test input, which felt like the first crack of warmth in the round.

Offer (October 20)

HR called with the verbal. The level was what I had targeted and the sponsorship sign-on was in the same offer. I had three days to think about it before written came through.

What TikTok MLE Actually Evaluates

After reflecting, four buckets:

1. Recommendation system literacy. Not just "I have used one," but "I can discuss candidate generation, ranking, rerank, and cold start with vocabulary precision."
2. Model math. Loss forms, gradient intuition, numerical considerations. Not exam-level derivations, but not surface-level either.
3. Pipeline thinking. End-to-end story from request to response, ownership boundaries, latency budgets.
4. Algorithm coding. Strictly LeetCode medium-to-hard, weighted toward DP, graphs, and heap problems. Not the loop's main signal, but an auto-fail if you bomb it.

Tips

1. Your resume projects are the actual test. TikTok MLE Round 2 and Round 3 both drilled my internship project for 20 minutes each. If you cannot defend every model choice, you should not list it. Strip your resume to projects you can discuss at the loss-function level.
2. Know Flash Attention, KV cache, and RoPE if you touch LLMs. Both senior interviewers used LLM vocabulary as a fluency check. If you say "I fine-tuned a transformer" and cannot discuss attention-memory tradeoffs, the conversation stops.
3. Have cold-start answers ready in both directions. New user with no history, new item with no interactions. Both are staple TikTok questions because they are daily reality on the app. Do not go in without a concrete answer for each.
4. Practice Hard DP under 15 minutes. The Head's coding problem was 15 minutes on the clock, and he was timing. Standard problems like partition-array, stock-with-k-transactions, and minimum-cost-subsequence are fair game. Being able to derive the recurrence in 2 minutes is the table stakes.
5. Sponsorship is in the offer conversation, not a separate negotiation. If you need sponsorship, say it to HR in Round 1. Do not surface it late. TikTok's MLE team specifically sponsors because their role has research weight, but not every team will.
6. The Head round decides the offer. If Round 2 is strong and Round 3 is weak, you get a reject. If Round 2 is average and Round 3 is strong, you can still get an offer. Do not budget equally.

TikTok MLE was the fastest loop I did this cycle and also the one I prepped hardest for the domain on. Recommendation systems literacy plus loss-function-level model defense is the bar. Everything else is cleanup.