Medusa Decoding Heads Quiz

Answers and explanations appear as you take the quiz.

1. What do the extra Medusa heads predict?
   • Multiple future tokens at different positions ahead
   • The model's training loss
   • Which GPU to use
   • The image caption for a token
2. Under which Medusa configuration does the paper describe lossless inference acceleration?
   • It compresses the model weights perfectly
   • Medusa-1 with a frozen backbone and exact verification against the original model
   • It never makes mistakes in grammar
   • It removes the need for a vocabulary
3. What structure are Medusa's candidate continuations arranged into for verification?
   • A flat list scored one by one
   • A hash table
   • A tree scored in a single pass with a tree attention mask
   • A recurrent loop over each token
4. What is a key advantage of Medusa over draft-model speculative decoding?
   • It requires two separate large models
   • It only works on images
   • It deletes the base model after use
   • It avoids needing a separate draft model by adding heads to the existing one
5. Roughly what speedup does Medusa typically report?
   • Around 2-3x
   • About 0.5x (slower)
   • Exactly 100x
   • No measurable change
6. What is each individual Medusa head, architecturally?
   • A full duplicate transformer
   • A small residual feed-forward (MLP) layer on top of the base hidden state
   • A convolutional image encoder
   • A separate tokenizer
7. What does Medusa accept from a verified candidate tree?
   • Every speculative branch regardless of verification
   • Only the shortest candidate token
   • The longest prefix consistent with the base model’s decoding result
   • A random leaf from the tree
8. How do Medusa’s prediction-head offsets differ?
   • Every head predicts the same next token
   • Each head predicts a different vocabulary
   • The heads predict tokens only after decoding is complete
   • Each head predicts a token at a different position ahead in the sequence