# Quiz with answers — Q/K/V attention: from projections to causal output ### Question 1 Solve the worked-case variant: Increase q’s first component from 2 to 2.4 without changing the keys. Recompute both scores divided by √4 and the attention weights. - A. Reduced trace: q=[2,1], keys k_Maya=[1,1], k_book=[0,2]. The real head has d_head=4; our vectors show only 2 coordinates (a truncated illustration), but the scaling keeps the true dimension. Raw scores: 3 and 2; division by √4=2 gives 1.5 and 1. Softmax ≈ [0.62,0.38]. Output mixes 62% of v_Maya and 38% of v_book. For an earlier position, the causal mask would remove every future key. - B. Attention maps alone do not prove a causal explanation of the model’s overall behavior. - C. Dot product q·k measures alignment. In “Maya put down the book… She picked it up”, one head can learn that the query for “She” aligns with the key for “Maya”. - D. Scaled scores move from [1.5,1] to [1.7,1]. Softmax weights move from about [0.622,0.378] to [0.668,0.332]: the Maya key receives more mass without becoming certain. **Answer: D.** Scaled scores move from [1.5,1] to [1.7,1]. Softmax weights move from about [0.622,0.378] to [0.668,0.332]: the Maya key receives more mass without becoming certain. The correct answer executes the requested change and gives a checkable result; the other texts do not close this calculation or trace. --- ### Question 2 Which causal order correctly connects the first three stages of “Q/K/V attention: from projections to causal output”? - A. Three learned projections → Query-key compatibility → Scaling - B. Scaling → Query-key compatibility → Three learned projections - C. Query-key compatibility → Three learned projections → Scaling - D. Three learned projections → Scaling → Query-key compatibility **Answer: A.** Three learned projections → Query-key compatibility → Scaling The chain follows the taught progression; reversing stages consumes a representation or state before it is produced. --- ### Question 3 If “Scaling” is removed, which diagnostic method is defensible? - A. Also change the data to amplify the difference. - B. Keep the same input, predict the first output that depends on “Scaling,” then compare the before/after trace. - C. Observe only the final output and invent the cause. - D. Conclude that the whole system fails before measuring. **Answer: B.** Keep the same input, predict the first output that depends on “Scaling,” then compare the before/after trace. One intervention and a prior prediction make the delta attributable to the removed mechanism. --- ### Question 4 Which verdict respects this session’s validity boundary? - A. The mechanism guarantees accuracy, speed, and stability for every workload. - B. One successful example proves the whole architecture is superior. - C. Attention maps alone do not prove a causal explanation of the model’s overall behavior. - D. The mechanism name alone is enough for a production choice. **Answer: C.** Attention maps alone do not prove a causal explanation of the model’s overall behavior. The correct answer bounds the conclusion; the others turn a local relation into a global guarantee. --- ### Question 5 Which evidence best matches the stated status of “Q/K/V attention: from projections to causal output”? - A. The route loads without an error. - B. Every learner opened the file. - C. The same result is assumed on every hardware target. - D. Established mechanisms; numerical simplifications are pedagogical. **Answer: D.** Established mechanisms; numerical simplifications are pedagogical. Product or mechanism evidence must remain attributed and measured; availability and completion do not prove value. --- ### Question 6 When should a simpler baseline be preferred to “Weighted value mixture”? - A. Never: the newest mechanism wins by default. - B. When a controlled test shows equivalent quality with lower memory, latency, or complexity. - C. As soon as one memory metric falls, regardless of quality. - D. As soon as the diagram contains fewer components. **Answer: B.** When a controlled test shows equivalent quality with lower memory, latency, or complexity. The choice depends on a measured trade-off on the real workload, not novelty or one isolated metric. --- ### Question 7 A learner gets the right result but cannot explain “Query-key compatibility.” Which remediation is most useful? - A. Rebuild the first missing transformation, label its inputs and outputs, then test a neighboring case. - B. Accept the answer because the final number is correct. - C. Provide the final result a second time. - D. Change several variables and ask for an intuition. **Answer: A.** Rebuild the first missing transformation, label its inputs and outputs, then test a neighboring case. The remediation targets the first causal break and then requires transfer instead of rewarding a guessed result. --- ### Question 8 Which submission actually demonstrates the outcome “Connect prefill, decode, and KV cache.”? - A. A list of terms without causal relations. - B. A screenshot without values or interpretation. - C. A trace with starting data, transformations, observed result, boundary, and next experiment. - D. A confident claim without a baseline or threshold. **Answer: C.** A trace with starting data, transformations, observed result, boundary, and next experiment. The correct submission makes the reasoning reproducible and the verdict revisable by future measurement.