Beam Search Quiz

Answers and explanations appear as you take the quiz.

1. What does the 'beam width' (k) control in beam search?
   • How many partial candidate sequences are kept at each decoding step
   • The maximum length of the generated sequence
   • The temperature applied to probabilities
   • The size of the model's vocabulary
2. How does beam search differ from greedy decoding?
   • It always picks the single highest-probability token
   • It keeps multiple high-scoring partial sequences instead of committing to one
   • It samples tokens at random
   • It ignores token probabilities entirely
3. Why is length normalization commonly applied in beam search?
   • To make generation faster
   • To increase the vocabulary size
   • Because summing log-probabilities biases scores toward shorter sequences
   • To add randomness to outputs
4. For which task is beam search especially well-suited?
   • Open-ended creative storytelling
   • Generating maximally diverse random text
   • Producing unpredictable poetry
   • Machine translation, where a faithful high-probability output is desired
5. What is a known downside of using a very large beam width for open-ended generation?
   • It can produce blander, repetitive, or degenerate text
   • It guarantees the most creative output
   • It removes the need for any probability model
   • It makes the model ignore the input
6. At each step, how does beam search decide which candidates to keep?
   • By the alphabetical order of tokens
   • By cumulative score (sum of log-probabilities), keeping the top k
   • By choosing the longest candidates
   • By random selection among all candidates
7. What problem does GNMT’s coverage penalty aim to reduce in translation?
   • Repeatedly encoding the same source vocabulary
   • The use of subword units
   • Output translations that fail to cover source words
   • The number of encoder layers
8. What does diverse beam search modify to encourage different candidate outputs?
   • It replaces the model probabilities with uniform random scores
   • It removes all but one candidate at the first step
   • It changes the training tokenizer after decoding
   • It penalizes similarity or overlap among candidates in different beam groups