Speculative Decoding with EAGLE Quiz

Answers and explanations appear as you take the quiz.

1. What is the core idea behind speculative decoding?
   • A cheap drafter proposes multiple tokens that the large model verifies in one parallel pass
   • Compressing the model's weights to half precision
   • Skipping layers of the network at random
   • Caching every possible output sentence in advance
2. How does EAGLE differ from earlier speculative decoding approaches?
   • It removes the verification step entirely
   • It drafts in the model's hidden feature space rather than directly at the token level
   • It uses a completely separate, larger model as the drafter
   • It only works for image generation
3. Why is speculative decoding considered lossless when implemented with exact verification?
   • Every draft token is accepted unconditionally
   • The drafter always predicts the target’s next token
   • The acceptance and correction procedure preserves the target model’s output distribution
   • The target model is never evaluated
4. What problem in EAGLE's feature drafting does feeding back the previous token's embedding help solve?
   • Memory fragmentation on the GPU
   • Tokenizer mismatch between languages
   • Overheating of the hardware
   • The uncertainty/ambiguity of predicting the next feature
5. What did EAGLE-2 add over the original EAGLE?
   • A dynamic draft tree that adapts how many candidates to propose
   • A fixed single-token draft
   • Removal of the LM head
   • Training on twice the data only
6. Why does the autoregressive bottleneck make plain LLM generation slow?
   • The model retrains itself after each token
   • Each new token needs a full forward pass through billions of parameters before the next can start
   • Tokens are stored on disk between steps
   • The GPU must reboot per token
7. After EAGLE predicts a hidden-state feature, which component turns that feature into candidate token probabilities?
   • The tokenizer’s vocabulary file alone
   • The optimizer’s momentum buffer
   • The target model’s language-model head
   • The KV-cache allocator
8. What change does EAGLE-3 make to the original feature-drafting constraint?
   • It replaces the target model with a tokenizer
   • It disables target verification
   • It requires a separate human drafter
   • It removes a feature-prediction constraint to improve scaling