Multi-Token Prediction Training Quiz

Answers and explanations appear as you take the quiz.

1. In the cited Meta study, what architecture predicts several future tokens from a training position?
   • A mandatory separate complete language model for every word
   • Independent output heads above a shared transformer trunk
   • One head that can target only the preceding token
   • A larger vocabulary with no additional prediction targets
2. Which model notably used a multi-token prediction objective in pretraining?
   • The original Word2Vec
   • A convolutional image classifier
   • An n-gram model
   • DeepSeek-V3
3. What additional supervision does multi-token training supply at a prefix position?
   • Loss terms for multiple future-token targets
   • Only a larger batch size with the same single target
   • A mandatory image-reconstruction target
   • No prediction loss at any position
4. How are auxiliary predictions used in the Meta paper’s self-speculative inference experiment?
   • They are accepted unconditionally without verification
   • They automatically retrain the model on every prompt
   • They propose tokens that are then checked by the decoding procedure
   • They encrypt every generated token
5. What happens to the auxiliary heads after training if you don't need speed-ups?
   • They must stay and double the model size
   • They can be discarded so deployment size is unchanged
   • They turn into attention layers
   • They become the tokenizer
6. Which evaluated task family showed notable gains in the Meta multi-token study?
   • Code generation benchmarks
   • A guarantee of perfect code compilation in every language
   • Only alphabetizing raw filenames
   • A required quantum-sensor prediction task
7. In a four-head offset design predicting from prefix position t, which token is the fourth head’s target?
   • The token at t−4
   • The token at t+1 for every head
   • Only the token at t itself
   • The token at t+4
8. What distinguishes DeepSeek-V3’s stated MTP architecture from the cited independent parallel-head implementation?
   • It predicts only tokens before the prefix
   • It requires every target to be an image pixel
   • Sequential auxiliary modules that preserve a causal chain across prediction depths
   • It removes the language-model output vocabulary