Unit 1: Foundations of LLMs and Agentic Systems - Practice Quiz

1 What is an intelligent agent?

2 Which type of agent acts only on the current percept using condition-action rules?

3 Which agent type improves its behavior through experience?

4 What distinguishes Agentic AI from a basic text-generation system?

5 Which capability is commonly associated with Agentic AI?

6 Which architecture forms the foundation of most modern large language models?

7 What information do positional encodings provide to a Transformer?

8 Which Transformer component converts token representations into contextual representations?

9 What does the attention mechanism help a language model do?

10 What does self-attention compare within a sequence?

11 Why are multiple attention heads used in a Transformer?

12 What is tokenization in an LLM pipeline?

13 Which tokenization approach represents text using pieces smaller than many complete words?

14 What is the usual objective of next-token prediction?

15 In masked language modeling, what does the model learn to predict?

16 What is the first main step in an agent's perception-action cycle?

17 After selecting an action, what does an agent normally do next?

18 What is a percept in agent-environment interaction?

19 What can an agent use to affect its environment?

20 Which organization developed the GPT family of language models?

21 A warehouse robot selects routes using a fixed set of condition-action rules and does not maintain an internal representation of previously observed obstacles. Which type of agent best describes it?

22 A delivery drone evaluates several safe routes and chooses the one that best balances travel time, energy use, and customer priority. Which agent type is most appropriate?

23 A language model is asked to organize a conference. It creates a task list, calls calendar and email tools, checks the responses, and revises its schedule. What most clearly makes this an agentic AI system?

24 An LLM agent repeatedly calls an expensive search tool even after it has enough evidence to answer. Which design addition would most directly reduce this behavior?

25 A transformer receives the same words in two different orders. Which component enables it to distinguish the sequences before attention is applied?

26 Why can transformer training usually process all tokens in a sequence more parallelly than a traditional recurrent neural network?

27 In a transformer block, what is the primary role of the position-wise feed-forward network?

28 In scaled dot-product attention, scores are divided by . What problem does this scaling mainly address?

29 In the sentence "The animal did not cross the street because it was tired," an attention head links "it" strongly to "animal." What capability does this illustrate?

30 Two attention heads in the same layer consistently focus on different relationships: one tracks nearby syntax, while another tracks long-range references. What is the main benefit?

31 A tokenizer encounters the rare word "microbioluminescence," which is absent from its vocabulary. Which approach is most likely to represent it using meaningful reusable pieces?

32 A model uses byte-level tokenization rather than a fixed word vocabulary. Which advantage is most relevant when processing multilingual text and unusual symbols?

33 A decoder-only model is trained on the sequence "Agents use tools" by predicting each token from only the tokens before it. Which objective is being used?

34 During pre-training, selected words in a sentence are hidden and the model predicts them using context on both sides. Which task does this describe?

35 A pre-trained LLM produces grammatically strong answers but does not reliably follow user instructions. Which additional training stage most directly targets this limitation?

36 A cleaning robot observes that a room remains dirty after one pass and then increases suction power for the next pass. Which sequence best represents this behavior?

37 A web-browsing agent submits a form and immediately assumes success without checking the resulting page. Which part of the perception-action cycle is missing?

38 A game agent can see only the area immediately around its character, while important objects may be elsewhere on the map. How should this environment be classified?

39 A trading simulation changes in response both to the agent's orders and to other participants' unpredictable decisions. Which description best captures the environment?

40 A team needs a model whose weights can be deployed on its own infrastructure and adapted for a specialized internal task. Which model family is commonly suited to this requirement?

41 An agent must select between two routes. Route X always yields utility 6. Route Y yields utility 20 with probability 0.4 and utility 0 otherwise. The agent knows these probabilities and consistently selects Route Y. Which classification best explains its behavior?

42 A language model generates a complete multi-step plan but cannot inspect results, call tools, revise actions, or preserve state. Which missing capability most directly prevents the system from being genuinely agentic?

43 In a deep pre-layer-normalized Transformer block, which update equation correctly places layer normalization while preserving an unnormalized residual stream?

44 A single attention head has , query , keys and , and scalar values , . With scaled dot-product attention and no mask, what is the output?

45 A BPE tokenizer is trained only on a fixed corpus and then applied to a new domain. Which claim about an unseen technical term is most accurate?

46 Two models use identical Transformer capacity. Model C is trained with causal language modeling, while Model M predicts independently selected masked tokens using bidirectional context. Which consequence follows directly from the objectives?

47 A warehouse agent observes shelf occupancy, acts, and then immediately plans again using its previous observation even though another robot may have moved items. What is the most precise defect in its perception-action cycle?

48 In a partially observable environment with known transition and observation models, what statistic is sufficient for optimal decision-making under the standard POMDP formulation?

49 Which architecture-objective mapping is correct for the original pre-training formulations of BERT, GPT-style models, and T5?

50 In the classical learning-agent decomposition, a component evaluates the agent's behavior against a performance standard and supplies feedback to the learning component. Which component performs this role?

51 An LLM-based system gains tool use, persistent memory, explicit goals, and iterative replanning. Which addition most fundamentally changes it from isolated text generation into an agent-environment system?

52 Suppose positional information is removed from a Transformer encoder and no attention masks depend on token position. What property does the resulting network have with respect to permutations of input tokens?

53 A decoder replaces standard multi-head attention with grouped-query attention while keeping the number of query heads unchanged. What is the principal inference-time effect?

54 A byte-level tokenizer and a word-level tokenizer have fixed vocabularies learned from English data. When processing arbitrary valid UTF-8 text from an unseen script, which comparative statement is correct?

55 A masked-language model masks each selected token independently, but the corruption token used during pre-training never occurs during downstream use. Which issue motivated BERT's mixed replacement strategy for selected positions?

56 An agent executes the first action of a five-step plan, observes an unexpected state that invalidates the remaining plan, but continues executing it unchanged. Which design principle would most directly correct this failure?

57 In a multi-agent environment, an agent's transition statistics change even though the physical rules remain fixed because other agents continually update their policies. From the focal agent's perspective, what challenge does this create?

58 Which statement correctly distinguishes the core architectures associated with the BERT, LLaMA, and T5 model families?

59 Compared with masked-language modeling, ELECTRA's replaced-token detection can obtain a training signal at more sequence positions. Why?

60 Rotary positional embeddings apply position-dependent rotations to query and key vector pairs. Which property explains their usefulness in attention scoring?