Mechanistic Interpretability Quiz

Answers and explanations appear as you take the quiz.

1. What is the central goal of mechanistic interpretability?
   • To reverse-engineer a network's internal computations into human-understandable algorithms
   • To make models train faster
   • To compress models for mobile devices
   • To rank input features by importance only
2. What does 'superposition' refer to in a neural network?
   • Stacking many models together
   • Representing more features than there are dimensions by using overlapping directions
   • Running a model on multiple GPUs
   • Averaging predictions from different runs
3. Sparse autoencoders are used in mechanistic interpretability primarily to...
   • Generate training data
   • Speed up inference
   • Disentangle polysemantic activations into cleaner, more interpretable features
   • Encrypt model weights
4. What are 'induction heads'?
   • Hardware accelerators for transformers
   • Output layers that generate text
   • A type of loss function
   • Attention heads that copy and continue repeated token patterns, supporting in-context learning
5. How do researchers confirm that a discovered circuit truly performs a hypothesized computation?
   • Through causal interventions like ablating or patching specific activations
   • By checking the model's training accuracy
   • By counting the number of parameters
   • By measuring inference latency
6. What is a cautious safety-related reason to study mechanistic interpretability?
   • It proves every hidden behavior can already be found
   • It may provide evidence about internal computations, while current methods are not a complete audit or direct steering guarantee
   • It guarantees a model will follow any requested correction
   • It replaces evaluations of model behavior in deployment
7. If a sparse autoencoder feature activates on examples described as “Golden Gate Bridge,” what does that observation establish?
   • The entire model’s computation is now explained
   • The feature is guaranteed to be a single neuron
   • A candidate feature correlates with those examples and needs further testing to establish its role
   • The feature causes every bridge-related output
8. Why test a candidate feature across varied prompts and examples?
   • Proof that the model has no other relevant circuits
   • A guarantee that the explanation generalizes to every prompt
   • A replacement for inspecting model behavior
   • To check whether the observed activation pattern generalizes beyond selected examples