Bytedance Data Scientist Interview Experience — Three Onsite Rounds on Multimodal Deployment and Dropout

Round 1

• Discussed how to deploy a multimodal model when compute and GPU memory are limited.
• Follow-up: given existing captions and embeddings, how to speed up video retrieval.
• Asked what overfitting is and how to mitigate it.
• Coding question: implement a MinStack that returns the minimum in O(1) time.

Round 2

• Discussed methods to mitigate overfitting in deep learning and how Dropout works.
• Compared different normalization methods and how they are handled at inference time.
• Discussed how reinforcement learning is used in LLM post-training (RLHF).
• Coding question: implement a MaxStack.
• Follow-up: how to compute the median in real time over a data stream, and how to modify the MaxStack implementation to support that.

Round 3

• Asked again to explain Dropout and why it keeps the distribution consistent.
• Coding question: given a binary tree, determine whether there is a path starting from any node, moving only upward, whose sum equals a target value.