TikTok New Grad Machine Learning Engineer Interview Experience — Three Technical Rounds, Pulled From the Pool

(Project questions omitted)

Round 1:
Our current business needs a multimodal large model to caption videos — how do you ensure deployment stays within compute and memory constraints? Follow-up: if you already have captions and embeddings, and a brand advertiser wants to quickly find relevant videos to watermark, how would you speed that up?
What is overfitting, and how do you mitigate it?
Coding: implement a MinStack that returns the current minimum in the stack in O(1).

Round 2:
How do you mitigate overfitting in machine learning? Narrow it down to deep learning specifically? Explain dropout?
Follow-up: explain the difference between two normalization methods? How do you handle the norm at deployment time?
Talk about common applications of RL in LLM post-training.
Coding: implement MaxStack. Follow-ups:
Given a large data stream, how would you return the real-time median?
If you wanted to implement this using MaxStack, what changes would you need to make?

Round 3:
Explain dropout. How do you make sure the distribution stays the same once you add dropout during training?
Coding: given a binary tree (all positive integers) and a target value, determine whether there's a single path (starting from some node and only moving upward) whose sum equals the target value.

P.S. This is the third time (fourth if you count my interviews back in China) I've interviewed for ByteDance MLE, and the first time I've made it through (T_T). This wasn't something I applied for myself either — ByteDance pulled me out of their candidate pool. Overall the interviewers came across well. There was a lot of overlap in the questions across the three technical rounds. Combining this with my past experience, it feels like for new grads, TikTok leans more on ML fundamentals, and they'll work in some knowledge questions based on your resume.