Constrained Decoding Quiz

Answers and explanations appear as you take the quiz.

1. What does constrained decoding fundamentally do at each generation step?
   • Masks out tokens that would violate the required structure before sampling
   • Retrains the model on structured examples
   • Lowers the temperature to zero
   • Adds more training data to the prompt
2. What can happen if a requested JSON Schema uses a feature unsupported by a structured-output implementation?
   • The model silently gains new training data
   • The request can be rejected with an error
   • All outputs become valid SQL
   • The schema is automatically converted to natural language
3. What technical challenge makes constrained decoding tricky to implement?
   • Models cannot output text at all without it
   • It requires a GPU with more than 100GB of memory
   • Subword tokens don't cleanly align with grammar symbols
   • Grammars cannot be written for JSON
4. Which of these is a real format used to specify constraints for decoding?
   • A spreadsheet of weights
   • A list of GPU temperatures
   • A random seed value
   • A JSON Schema or context-free grammar
5. At what point in the pipeline is the constraint mask typically applied?
   • To the logits, before the softmax/sampling step
   • Before tokenization of the prompt
   • After the text is already printed to the user
   • Only during model training
6. Compared to simply asking a model in the prompt to 'output valid JSON', constrained decoding offers what advantage?
   • It uses less electricity
   • It provides a hard guarantee rather than relying on the model's cooperation
   • It makes the model more creative
   • It removes the need for a tokenizer
7. What does PICARD do with a candidate token that would make a partial SQL query unparsable?
   • Accepts it and repairs the query after generation
   • Retrains the language model immediately
   • Rejects that token during incremental parsing before it is emitted
   • Deletes the user’s database schema
8. What does a schema-conforming output guarantee by itself?
   • Every value is true in the external world
   • The model used the right source documents
   • No refusal or incomplete response can occur
   • The output matches the supported structure, not that its values are factually correct