Function Vectors and Task Representations Quiz

Answers and explanations appear as you take the quiz.

1. What does a function vector represent in the studied language models?
   • A compact activation-space representation associated with a demonstrated task
   • The model’s full weight matrix
   • A single token embedding in every setting
   • The optimizer’s learning rate
2. How is a function vector typically constructed?
   • By summing all word embeddings in the vocabulary
   • By averaging the outputs of specific task-carrying attention heads over example prompts
   • By fine-tuning the model on one example
   • By reading the final softmax probabilities
3. What did the function-vector paper observe after injecting a vector into a zero-shot prompt?
   • The model always crashed
   • It guaranteed correct answers on any task
   • It could trigger the associated task in the tested settings
   • It permanently changed the model weights
4. Which analysis technique underpins finding the heads that carry a function vector?
   • Gradient descent on the loss
   • Beam search
   • Dropout regularization
   • Causal mediation analysis
5. What do function-vector experiments provide evidence for?
   • Some models form reusable internal representations associated with tasks
   • All language models only memorize text
   • Every model has one universal task vector
   • Tokenization alone determines every behavior
6. Why are function vectors interesting to steering research?
   • They guarantee safe or correct behavior
   • They provide an activation-level intervention that can influence task behavior in tested settings
   • They eliminate prompts for every application
   • They make all model decisions fully transparent
7. At what part of the transformer did the paper report strong function-vector effects?
   • Only the input embedding table
   • Only the final output logits
   • Middle layers in the tested models
   • The tokenizer vocabulary
8. What happened when researchers summed some function vectors?
   • The vectors always canceled to zero
   • The model’s weights were permanently updated
   • Every possible task was composed perfectly
   • Some combinations triggered new complex tasks in the tested experiments