Speculative Edits for Code Models Quiz

Answers and explanations appear as you take the quiz.

1. What key observation makes speculative edits effective?
   • Most of an edited file's tokens are usually identical to the original
   • AI models cannot read code at all
   • Files must always be regenerated from scratch
   • Code has no predictable structure
2. Where does the 'draft' come from in speculative edits, compared to ordinary speculative decoding?
   • From a separate small draft model
   • From the existing source code being edited
   • From random sampling
   • From a human reviewer
3. How does standard autoregressive decoding produce tokens?
   • Many tokens in parallel per pass
   • All tokens at once with no model
   • One token per forward pass
   • Only on a GPU cluster
4. In speculative methods, what happens to a run of proposed tokens that the model verifies as correct?
   • They are discarded and regenerated
   • They are sent to a human for approval
   • They cause an error
   • They are accepted in roughly a single forward pass
5. With speculative edits, the generation cost scales most closely with what?
   • The size of the actual edit (the divergent parts)
   • The total size of the file
   • The number of files in the repo
   • The model's parameter count
6. Which scenario benefits MOST from speculative edits?
   • Writing a brand-new file from an empty buffer
   • Renaming a symbol across a large mostly-unchanged file
   • Answering a general trivia question
   • Generating a short standalone poem
7. A code edit changes only one short function in a large file. Where is speculative generation work concentrated?
   • Across every unchanged token in the file
   • In retraining the code model
   • In the divergent edited span that does not match the existing source draft
   • In regenerating the file’s tokenizer vocabulary
8. When speculative verification finds a mismatch with the target model, what should the edit system do?
   • Accept every remaining draft token
   • Discard the entire model checkpoint
   • Replace the target model with the draft
   • Stop accepting that draft prefix and generate the corrected span