I recently had my first online interview at Amazon’s Bengaluru office recently in May 2025. I have 4 years of industry experience as a Data Scientist in a mid sized company. The round was focused on theoretical understanding of large language models (LLMs), transformer internals, and general ML/statistics concepts.
Description:
I was asked to explain the concept of QKV (Query, Key, Value) in the attention mechanism, and specifically how it’s implemented and used differently in BERT vs GPT.
Follow-up:
Why does GPT only use causal masking? What architectural changes exist in decoder-only vs encoder-decoder models?
⸻
Description:
They wanted to check my statistical intuition. I was asked to define a point estimator, and explain the bias-variance tradeoff with relevant examples.
Follow-up:
What would you prefer in a low-data regime: high bias or high variance model?
⸻
Description:
What are the key sources of stochastic behavior in LLMs during inference?
Follow-up:
How does temperature sampling and top-k/top-p affect generation?
⸻
Description:
This was more of a conceptual discussion. I was asked:
“How do LLMs generalize well with very little task-specific data during fine-tuning or prompting?”
Follow-up:
Why doesn’t the classical bias-variance limitation seem to apply here?
Further round scheduled which would focus on the coding aspects alongwith deep dive into resume projects.