Contrastive Decoding Quiz

Answers and explanations appear as you take the quiz.

1. In contrastive decoding, what role does the small 'amateur' model play?
   • Its log-probabilities are subtracted to suppress shared failure modes
   • It generates the final output text directly
   • It is fine-tuned on the expert's mistakes
   • It replaces the expert when the expert is uncertain
2. Why does subtracting the amateur's probabilities reduce repetition and bland output?
   • The amateur is trained never to repeat words
   • Generic and repetitive tokens are favored by both models, so the contrast cancels them
   • The amateur deletes duplicate tokens after generation
   • Repetition is impossible once two models vote
3. What does the adaptive plausibility constraint do in contrastive decoding?
   • It guarantees every remaining continuation is coherent and true
   • It replaces the expert distribution with the amateur distribution
   • It filters out candidates whose expert probability is too low relative to the expert’s top token
   • It forces every token to have equal probability
4. Roughly what is the scoring formula in contrastive decoding?
   • log p_amateur minus log p_expert
   • p_expert times p_amateur
   • The average of the two probabilities
   • log p_expert minus a coefficient times log p_amateur
5. A key practical advantage of contrastive decoding is that it:
   • Works at inference time with no extra training
   • Requires retraining both models jointly
   • Only works on classification tasks
   • Eliminates the need for a large model
6. Which later method extends the contrastive idea to a single model's own layers?
   • Beam search
   • DoLa, which contrasts early versus late layers
   • Greedy decoding
   • Temperature scaling
7. The expert assigns probability 0.4 to its top token and the plausibility threshold is 0.1 of that value. What is the minimum probability for a candidate to remain eligible?
   • 0.4
   • 0.1
   • 0.04
   • 0.004
8. Why can contrastive decoding require more inference work than using the expert model alone?
   • It must retrain both models for every prompt
   • It replaces the vocabulary with image patches
   • It runs only the amateur model
   • It scores candidates with both an expert and an amateur model at inference time