Unit 2: Exploring Large Language Models (LLMs) - Practice Quiz

1 What is a Large Language Model (LLM)?

2 The word "Large" in Large Language Model mainly refers to which characteristic?

3 Which architecture is most commonly used as the foundation for modern LLMs?

4 What is a token in the context of LLMs?

5 What does an LLM primarily learn to predict during training?

6 Which of the following is an example of a Large Language Model?

7 Which of the following is a common use case for LLMs?

8 An LLM that answers customer questions automatically is being used as a:

9 Which task involves an LLM converting text from one language to another?

10 Generating computer code from a natural language description is known as:

11 Analyzing whether a product review is positive or negative is an example of:

12 Which of the following is NOT a typical use case for an LLM?

13 What is the main goal of prompt tuning?

14 In LLMs, a prompt refers to:

15 How does prompt tuning differ from full model fine-tuning?

16 Giving an LLM a few examples inside the prompt to guide its answer is called:

17 A key benefit of prompt tuning is that it:

18 Which Google platform provides tools for building and deploying generative AI applications on Google Cloud?

19 What is the name of Google's family of large language models?

20 Which Google tool lets developers quickly prototype prompts and test Gemini models?

21 A team notices that their LLM can perform a translation task it was never explicitly trained on, simply by being shown a few examples in the prompt. Which characteristic of LLMs does this demonstrate?

22 Why are modern LLMs described as "foundation models" rather than task-specific models?

23 Which architecture is the primary foundation of most contemporary LLMs, enabling them to weigh the relevance of different words in a sequence?

24 An LLM is asked to complete the sentence "The capital of France is ___". Fundamentally, what is the model doing to produce the answer?

25 The term "large" in Large Language Models most directly refers to which two factors?

26 Which phenomenon describes an LLM producing fluent but factually incorrect or fabricated information?

27 A company wants to automatically condense long legal documents into short overviews for busy executives. Which LLM use case fits this need?

28 An online retailer wants to route customer emails automatically as "complaint", "inquiry", or "feedback". Which LLM use case is most appropriate?

29 A developer uses an LLM-powered tool that suggests function implementations as they type. Which use case does this represent?

30 A support team wants an LLM to answer questions using their internal knowledge base so responses stay accurate and up to date. Which approach best addresses this?

31 Which scenario is generally the LEAST suitable standalone use case for an LLM without additional tools?

32 A media firm wants to convert English articles into Spanish and French for global readers. Which LLM use case applies?

33 How does prompt tuning primarily differ from full fine-tuning of an LLM?

34 A team has a very large model and limited GPU memory. Why might prompt tuning be preferred over full fine-tuning?

35 In prompt tuning, what are "soft prompts"?

36 Which statement best contrasts prompt engineering with prompt tuning?

37 Prompt tuning is best categorized as which type of technique?

38 A developer wants a unified Google Cloud platform to build, tune, and deploy generative AI models with access to foundation models. Which tool should they use?

39 Within Vertex AI, which environment lets developers quickly test prompts and compare model responses before writing production code?

40 A team wants to integrate Google's Gemini models into their app using a simple developer API and lightweight SDK. Which offering is designed for this?

41 A researcher observes that a decoder-only transformer LLM shows improved sample efficiency and emergent few-shot capabilities only after crossing a certain parameter count. Which principle best explains why simply increasing depth without scaling data and compute proportionally often fails to produce these gains?

42 In an autoregressive LLM, the training objective maximizes . Which consequence of this factorization most directly explains why such models can suffer from exposure bias at inference?

43 Two LLMs have identical architectures and parameter counts, but Model A uses a byte-pair encoding (BPE) tokenizer with a 32k vocabulary while Model B uses a 256k vocabulary. Which trade-off is most accurately attributed to Model B?

44 An LLM's self-attention has time complexity for sequence length and hidden dimension . A team wants to process documents 4× longer without a 16× compute blowup. Which approach fundamentally changes the complexity class rather than just optimizing constants?

45 Why do LLMs using absolute learned positional embeddings often generalize poorly to sequences longer than those seen during training, whereas rotary (RoPE) or relative embeddings handle length extrapolation better?

46 An enterprise deploys an LLM for retrieval-augmented generation (RAG) over proprietary documents. Users report the model still fabricates facts even with correct passages retrieved. Which root cause is the most likely and directly addressable within the RAG pipeline?

47 A company must choose between fine-tuning an LLM on 500 labeled examples versus using few-shot in-context learning for a specialized classification task. Which consideration most strongly favors fine-tuning in this scenario?

48 For a code-generation assistant, engineers notice that increasing the sampling temperature from to raises the diversity of solutions but lowers pass rates on unit tests. Which use-case-driven configuration best balances correctness with exploring alternatives?

49 A summarization pipeline processes documents far longer than the model's context window. Which strategy preserves global coherence better than naive independent chunk summarization?

50 An LLM-powered agent chains tool calls (search, calculator, database). Reliability drops sharply as the number of sequential steps grows. If each step succeeds independently with probability , what is the approximate end-to-end success probability over 5 steps, and what does this imply?

51 In soft prompt tuning, a set of continuous embedding vectors is prepended to the input and optimized while the base model weights stay frozen. Which statement most accurately distinguishes this from discrete prompt engineering?

52 Research shows prompt tuning becomes competitive with full fine-tuning primarily as model scale increases. Which explanation best accounts for this scale-dependent effectiveness?

53 A team compares prompt tuning, prefix tuning, and LoRA for adapting one base model to many tasks. Which property is shared by all three as parameter-efficient fine-tuning (PEFT) methods?

54 If a soft prompt of length tokens is prepended to an input of length , how does this affect the self-attention cost during a forward pass, and why is it still efficient overall?

55 A practitioner finds their prompt-tuned model performs well in-distribution but degrades sharply on inputs stylistically different from training data. Which limitation of prompt tuning best explains this brittleness?

56 A developer needs to build, tune, and deploy generative models with managed infrastructure, MLOps, and access to foundation models like Gemini. Which Google offering is the most appropriate end-to-end platform?

57 A team wants to rapidly prototype and iterate on prompts for Gemini, test parameters like temperature, and then export code to integrate into an application. Which Google tool is designed for this prompt-iteration workflow?

58 An enterprise on Vertex AI wants to ground Gemini responses in its own frequently-updated document corpus without retraining the model each time. Which capability most directly supports this requirement?

59 When adapting a Gemini model on Vertex AI with limited labeled data and constrained budget, which tuning approach aligns with parameter-efficient adaptation rather than full fine-tuning?

60 A regulated organization must call Gemini from production code with governance controls: IAM permissions, VPC service controls, audit logging, and data residency. Which access path best satisfies these enterprise requirements?