Interpretability interview questions
10 questions across 6 sections, in the order you would learn them. Every answer here is one that works — but an answer you have read is not an answer you can give, so use Test yourself and say it out loud first.
Working out what the network is doing
Read the section that teaches thisIf this comes up in an interview
What is mechanistic interpretability?
Reverse-engineering the computation a network performs, identifying circuits of components that implement a specific behaviour. More ambitious than attribution or probing, which say where the model looked or what information is present, not what it computed.
What is superposition and why does it make interpretability hard?
Models represent more features than they have dimensions by packing them into overlapping directions, so one neuron responds to several unrelated things. That is why reading individual neurons rarely works, and why sparse autoencoders are used to separate features.
What can interpretability actually establish today?
Specific circuits and specific behaviours have been explained, mostly in smaller models. Explaining an arbitrary output from a frontier model is not yet possible. It is real progress on a problem that is far from solved.
The techniques, and what each establishes
Read the section that teaches thisIf this comes up in an interview
What is the difference between probing and activation patching?
Probing shows information is present in a representation, which is correlational and compatible with the model never using it. Patching replaces an activation from one run into another and checks whether the output changes, which establishes that the output causally depends on it.
What is superposition and what is the response to it?
Models represent more features than they have dimensions by packing them into overlapping directions, so individual neurons respond to several unrelated things. Sparse autoencoders, trained with many hidden units and a sparsity penalty, attempt to separate those features back out.
Why is attribution not an explanation?
It shows which inputs influenced the output, which is consistent with correct reasoning and with pattern matching. It is evidence about where the model attended rather than about what it computed.
Looking inside a transformer with TransformerLens
Read the section that teaches thisIf this comes up in an interview
What is the residual stream, and why is it a useful way to think about a transformer?
The per-position vector that runs from embedding to output, which each attention and MLP layer reads from and adds to. Because contributions are added rather than overwritten, you can attribute the final output to individual components and ask what each one contributed.
Then they push further
What does the logit lens show, and what is its limitation?
It decodes intermediate residual states with the final unembedding to show the model's best guess at each layer. It assumes intermediate states live in the same space as the final one, which holds only roughly, especially in early layers.
How would you find an induction head?
Feed a random token sequence repeated twice and look for heads that, in the second half, attend strongly to the position just after the previous occurrence of the current token. Then confirm causally by ablating them and watching loss on the repeated half rise.
What gets marked up. Marked up for the additive view of the residual stream and for confirming a found circuit causally, not just by looking at attention.
Activation patching, in code
Read the section that teaches thisIf this comes up in an interview
Walk me through an activation patching experiment.
Choose a behaviour and a clean prompt, make a corrupted prompt that changes the answer by a minimal edit, pick a metric such as the logit difference between the two answers, then copy activations from the clean run into the corrupted run one location at a time and measure how much of the clean behaviour each restores.
Then they push further
Why is this stronger evidence than probing?
A probe shows information can be decoded from a location. Patching intervenes: if replacing that activation changes the output, the model's computation actually depends on it there.
What could make a patching result misleading?
An unrepresentative corruption, a metric that saturates, one prompt standing in for a whole behaviour, and backup components that take over when one is changed. Replicate across many prompt pairs and corruptions.
What gets marked up. Strong answers explain clean, corrupted and metric, stress the causal nature, and volunteer the limitations.
Sparse autoencoders, hands on
Read the section that teaches thisIf this comes up in an interview
How would you check whether an SAE feature really means what its label says?
Read activations across the whole range, not only the top examples, test predictions on new text where the label says it should and should not fire, and intervene by adding or removing the feature direction to see whether the model's behaviour changes as the label predicts.
Doing interpretability research
Read the section that teaches thisIf this comes up in an interview
How would you start an interpretability project?
Replicate a known result to validate the tooling, then pick one narrow behaviour in a small model, define a clean and corrupted setup, and run patching with controls against random components across many prompts, deciding in advance what would falsify the hypothesis.
These are the questions for one topic. The whole question bank covers every topic on the site.