# Quiz with answers — Kimi Delta Attention: a bounded case study ### Question 1 Solve the worked-case variant: Increase the second gate channel from [0.9,0.2] to [0.9,0.24]. Compare the trace retained in that channel after one step with all other values fixed. - A. For an initial trace of 1, the second channel retains 0.24 instead of 0.20, which is 20% more. The first channel stays at 0.9, so the vector gate changes memory channel by channel. - B. For vector gate [0.9,0.2], the first channel retains almost all its trace while the second forgets quickly. The same delta correction can therefore have different temporal effects by channel. - C. Exact Kimi K3 specifications are not claimed as verified in this lesson; the vendor technical report (Kimi Linear, arXiv:2510.26692, to verify) is cited without independent replication. - D. Replace the scalar with a vector: a = [0.9, 0.2], S′ = diag(a)·S. Channel 1 retains — half-life ln(0.5)/ln(0.9) ≈ 6.6 tokens — while channel 2 flushes (≈ 0.43): a 15× ratio inside the same state. **Answer: A.** For an initial trace of 1, the second channel retains 0.24 instead of 0.20, which is 20% more. The first channel stays at 0.9, so the vector gate changes memory channel by channel. 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 “Kimi Delta Attention: a bounded case study”? - A. Delta correction → Per-channel decay → Established mechanism - B. Established mechanism → Per-channel decay → Delta correction - C. Per-channel decay → Established mechanism → Delta correction - D. Established mechanism → Delta correction → Per-channel decay **Answer: B.** Established mechanism → Per-channel decay → Delta correction The chain follows the taught progression; reversing stages consumes a representation or state before it is produced. --- ### Question 3 If “Delta correction” is removed, which diagnostic method is defensible? - A. Also change the data to amplify the difference. - B. Observe only the final output and invent the cause. - C. Keep the same input, predict the first output that depends on “Delta correction,” then compare the before/after trace. - D. Conclude that the whole system fails before measuring. **Answer: C.** Keep the same input, predict the first output that depends on “Delta correction,” 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. The mechanism name alone is enough for a production choice. - D. Exact Kimi K3 specifications are not claimed as verified in this lesson; the vendor technical report (Kimi Linear, arXiv:2510.26692, to verify) is cited without independent replication. **Answer: D.** Exact Kimi K3 specifications are not claimed as verified in this lesson; the vendor technical report (Kimi Linear, arXiv:2510.26692, to verify) is cited without independent replication. 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 “Kimi Delta Attention: a bounded case study”? - A. Mixed: established mechanisms + source-reported Kimi K3-style choices. - B. The route loads without an error. - C. Every learner opened the file. - D. The same result is assumed on every hardware target. **Answer: A.** Mixed: established mechanisms + source-reported Kimi K3-style choices. 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 “Source-reported”? - 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 “Per-channel decay.” Which remediation is most useful? - A. Accept the answer because the final number is correct. - B. Provide the final result a second time. - C. Rebuild the first missing transformation, label its inputs and outputs, then test a neighboring case. - D. Change several variables and ask for an intuition. **Answer: C.** 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 “Classify claims by evidence strength.”? - A. A list of terms without causal relations. - B. A screenshot without values or interpretation. - C. A confident claim without a baseline or threshold. - D. A trace with starting data, transformations, observed result, boundary, and next experiment. **Answer: D.** 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.