How many Google questions are in this list?

59 questions are published in this slice right now (out of 100 requested). More land as our curation team publishes them.

Are these real AI interview questions?

Yes. Every question is sourced from real AI interview reports across companies like OpenAI, Anthropic, Google, Microsoft, and Cohere, then reviewed by senior AI engineers for accuracy.

Do these questions come with detailed answers?

Yes. Every entry links to a full answer with explanations, hints, key tradeoffs, and recommended follow-up questions. Most also include difficulty ratings and community quality scores.

How is the ranking calculated?

Questions are sorted by community quality score, with rating count as the tiebreaker, then difficulty (easier first). Quality scores come from interview candidates rating each question after working through it.

Where can I find more Google questions?

Beyond this Top 59 list, the full Google library on Zenaique includes every published question, filterable by format, difficulty, role, and experience level.

Curated · Google

Top 100 Google Interview Questions

Here are the top 59 AI interview questions tagged for Google interviews, sorted by community quality. Each entry links to a detailed answer with explanations, hints, and source-grade follow-ups. See all at this company →

Sorted by community quality · refreshed as new questions are published.

Clear filter

Resize10 15 20 25 50 100

01
What is RLHF, and why is it used after pretraining?
Multiple ChoiceEasy·Qual 4.7
02
Explain why transformers replaced RNNs for language modeling.
Short AnswerEasy·Qual 4.6
03
PPO vs DPO, what's the practical difference?
FlashcardMedium·Qual 4.6
04
BPE grows the vocabulary by adding merges. Unigram does the opposite. Explain the difference.
FlashcardEasy·Qual 4.0
05
Name a 2026 task where an encoder-decoder transformer still beats a decoder-only LLM
Short AnswerEasy·Qual 4.0
06
Pick the 2026 task where an encoder-only transformer still beats a decoder-only LLM
Multiple ChoiceEasy·Qual 4.0
07
Pair each transformer family with its canonical pretraining objective
Match PairsEasy·Qual 4.0
08
RLOO's name hides its baseline trick: what does 'Leave-One-Out' literally compute?
Multiple ChoiceMedium·Qual 4.0
09
Why are small FT datasets especially vulnerable to membership inference?
Multiple ChoiceMedium·Qual 4.0
10
Design a fine-tune that teaches a model to refuse when the retrieved context lacks the answer.
Short AnswerMedium·Qual 4.0
11
Sequence the forward passes and loss steps inside a single DPO training step
Order StepsMedium·Qual 4.0
12
Why is reporting only mean eval score dangerous for LLM quality tracking?
Multiple ChoiceMedium·Qual 4.0
13
Which Unicode normalization does BERT's tokenizer apply, and how does it differ from NFC?
Multiple ChoiceMedium·Qual 4.0
14
How does WordPiece decide which pair to merge, compared to BPE?
Multiple ChoiceMedium·Qual 4.0
15
Tulu-style instruction tuning trades volume for diversity: which axes of diversity matter most?
Multi-selectMedium·Qual 4.0
16
BPE-dropout is applied at which stage(s) of the LLM lifecycle?
Multiple ChoiceMedium·Qual 4.0
17
Which tokenizer can learn merges that span whitespace boundaries by design?
Multiple ChoiceMedium·Qual 4.0
18
Which statement correctly describes the Unigram Language Model tokenizer's training procedure?
Multiple ChoiceMedium·Qual 4.0
19
Order the steps in building a 5k-pair DPO dataset on helpfulness.
Order StepsMedium·Qual 4.0
20
Which of the following are failure modes unique to browser agents that do not occur with structured API tool calling?
Multi-selectMedium·Qual 4.0
21
What is the primary validity threat to Chatbot Arena ELO rankings as a measure of general model quality?
Multiple ChoiceMedium·Qual 4.0
22
Explain why word-level and character-level tokenization both fail for modern LLMs, and how subword tokenization resolves each failure mode.
Short AnswerMedium·Qual 4.0
23
Where do SFT and DPO sit in the classical RLHF pipeline?
Short AnswerMedium·Qual 4.0
24
For serving a Llama model in production with HuggingFace transformers, which tokenizer library is recommended?
Multiple ChoiceMedium·Qual 4.0
25
What does the router inside a Mixture-of-Experts block actually do, per token?
FlashcardMedium·Qual 4.0
26
Match each preference-tuning variant to its design choice on noisy labels
Match PairsMedium·Qual 4.0
27
Why has GRPO eaten PPO's lunch in 2025-2026 reasoning post-training?
Short AnswerMedium·Qual 4.0
28
When would you pick KTO over DPO?
Short AnswerMedium·Qual 4.0
29
You tell the model 'respond in exactly 3 bullet points' and it gives you 5. Which benchmark specifically targets this class of failure?
Multiple ChoiceMedium·Qual 4.0
30
DPO vs RLHF: pick the most accurate operational difference
Multiple ChoiceMedium·Qual 4.0
31
What are the strengths and weaknesses of MMLU as an LLM benchmark?
Multiple ChoiceMedium·Qual 4.0
32
Which design choices from the original 'Attention Is All You Need' are no longer standard?
Multi-selectMedium·Qual 4.0
33
Membership inference on fine-tuned models. Why are small FT datasets vulnerable?
Short AnswerHard·Qual 4.0
34
Analyze MMLU's strengths and failure modes as an LLM benchmark
Short AnswerHard·Qual 4.0
35
Spend a fixed parameter budget: more layers or wider d_model, and what does the evidence say?
Short AnswerHard·Qual 4.0
36
How does weighting BPE merges by document frequency instead of raw token frequency change the resulting vocabulary?
Short AnswerHard·Qual 4.0
37
Naive safety FT failed: model now refuses 'how does anesthesia work?'. Best fix?
Multiple ChoiceHard·Qual 4.0
38
Why does fine-tuning on company docs often raise hallucination instead of lowering it?
Multiple ChoiceHard·Qual 4.0
39
DPO failure modes: large β, small β, noisy preferences
Short AnswerHard·Qual 4.0
40
You are building a 7B model for 25 languages including Hindi, Arabic, and Swahili. Design a fair tokenizer and name what it costs.
Short AnswerHard·Qual 4.0
41
Select the correct comparisons of an MoE transformer versus a dense transformer at equal active parameters
Multi-selectHard·Qual 4.0
42
Design a slice-level eval strategy for an LLM system where mean score is insufficient
Short AnswerHard·Qual 4.0
43
Browser agents drive a real UI instead of calling text APIs. What extra failure modes show up?
Short AnswerHard·Qual 4.0
44
DPO: what's the closed-form trick that lets it skip the reward model?
Short AnswerHard·Qual 4.0
45
Compare MHA, MQA, GQA, MLA, what production tradeoff are they all addressing, and which models use each?
Short AnswerHard·Qual 4.0
46
Why does SentencePiece prepend a special space symbol (▁) to tokens, and what would break without it?
Short AnswerHard·Qual 4.0
47
Which methods are practical approaches for detecting train-eval data leakage in LLM benchmarks?
Multi-selectHard·Qual 4.0
48
Over-refusal: what it is and how naive safety FT causes it
Short AnswerHard·Qual 4.0
49
Spot the failure mode: a tokenizer trained with NFC normalization is deployed with an NFKC preprocessing pipeline.
Spot the ErrorHard·Qual 4.0
50
Define 'fertility' as a tokenizer metric and explain why high fertility for a language is both a fairness and a model performance problem.
Short AnswerHard·Qual 4.0
51
How does vocabulary size affect embedding table memory footprint at bfloat16 precision, and what are the tradeoffs of very small vs. very large vocabularies?
Short AnswerHard·Qual 4.0
52
Explain the Unigram LM tokenization algorithm and how its training differs fundamentally from BPE. What does stochastic tokenization enable?
Short AnswerHard·Qual 4.0
53
Chinchilla's lesson: how did it reshape architecture and training choices after 2022?
Short AnswerHard·Qual 4.0
54
SimPO: what does it drop vs DPO, and what's the cost win?
Short AnswerHard·Qual 4.0
55
ORPO vs DPO: which architectural difference enables 'single-stage' training?
Multiple ChoiceHard·Qual 4.0
56
ORPO: what does it combine into one stage, and why is that attractive?
Short AnswerHard·Qual 4.0
57
What are the practical differences between the sentencepiece library and HuggingFace tokenizers for serving a Llama model, and which is recommended?
Short AnswerHard·Qual 4.0
58
Pair each 2017 transformer ingredient with its 2026 production replacement
Match PairsHard·Qual 4.0
59
Why can fine-tuning on domain facts INCREASE hallucination?
Short AnswerHard·Qual 4.0

59 of 100 requested: only 59 published questions tagged for Google so far.

Related top picks

By experience level

Fresher Mid-Level Senior Career Switcher

Other companies

Top 100 Anthropic Interview Questions→Top 100 OpenAI Interview Questions→Top 100 Databricks Interview Questions→

By topic

Top 100 Inference Optimization Interview Questions→Top 100 Fine-Tuning Interview Questions→Top 100 LLM Evaluation Interview Questions→