Activation Patching: Intervention Experiments

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-safety2-activation-patching-creators

1. What is the core idea behind "Activation Patching: Intervention Experiments"?

   1. Correlation is not causation, even inside a neural network. Activation patching is the interpretability equivalent of a controlled experiment — swap one component and see what changes.
   2. autonomous agent
   3. Whistleblower protections for lab employees
   4. algorithmic registration

2. Which term best describes a foundational idea in "Activation Patching: Intervention Experiments"?

   1. path patching
   2. activation patching
   3. causal tracing
   4. ablation

3. A learner studying Activation Patching: Intervention Experiments would need to understand which concept?

   1. activation patching
   2. causal tracing
   3. path patching
   4. ablation

4. Which of these is directly relevant to Activation Patching: Intervention Experiments?

   1. activation patching
   2. path patching
   3. ablation
   4. causal tracing

5. Which of the following is a key point about Activation Patching: Intervention Experiments?

   1. Run the model on a clean input — get the correct answer
   2. Run the model on a corrupted input (e.g., a different name, a shuffled prompt) — get the wrong answe…
   3. Save activations from the clean run at every position and layer
   4. Re-run the corrupted input, but patch in a clean activation at one specific site

6. Which of these does NOT belong in a discussion of Activation Patching: Intervention Experiments?

   1. autonomous agent
   2. Run the model on a corrupted input (e.g., a different name, a shuffled prompt) — get the wrong answe…
   3. Run the model on a clean input — get the correct answer
   4. Save activations from the clean run at every position and layer

7. Which statement is accurate regarding Activation Patching: Intervention Experiments?

   1. Works without retraining
   2. Scales to any architecture with a residual stream
   3. Produces causal, not correlational, evidence
   4. Can identify redundant circuits that do the same work

8. Which of these does NOT belong in a discussion of Activation Patching: Intervention Experiments?

   1. Scales to any architecture with a residual stream
   2. Produces causal, not correlational, evidence
   3. Works without retraining
   4. autonomous agent

9. What is the key insight about "Path patching" in the context of Activation Patching: Intervention Experiments?

   1. A refinement: patch not just a component but a specific pathway between two components.
   2. autonomous agent
   3. Whistleblower protections for lab employees
   4. algorithmic registration

10. What is the key insight about "Off-distribution activations" in the context of Activation Patching: Intervention Experiments?

    1. autonomous agent
    2. When you patch an activation from one input into another, the receiving layer sees something it would never see in norma…
    3. Whistleblower protections for lab employees
    4. algorithmic registration

11. Which statement accurately describes an aspect of Activation Patching: Intervention Experiments?

    1. autonomous agent
    2. Whistleblower protections for lab employees
    3. A probe tells you a model represents something. A circuit diagram is a hypothesis about how.
    4. algorithmic registration

12. What does working with Activation Patching: Intervention Experiments typically involve?

    1. autonomous agent
    2. Whistleblower protections for lab employees
    3. algorithmic registration
    4. The big idea: patching is how you separate 'the model represents this' from 'the model uses this.

13. Which best describes the scope of "Activation Patching: Intervention Experiments"?

    1. It focuses on Correlation is not causation, even inside a neural network. Activation patching is the interpretabil
    2. It is unrelated to ethics workflows
    3. It applies only to the opposite beginner tier
    4. It was deprecated in 2024 and no longer relevant

14. Which section heading best belongs in a lesson about Activation Patching: Intervention Experiments?

    1. autonomous agent
    2. The standard procedure
    3. Whistleblower protections for lab employees
    4. algorithmic registration

15. Which section heading best belongs in a lesson about Activation Patching: Intervention Experiments?

    1. autonomous agent
    2. Whistleblower protections for lab employees
    3. Why it is gold-standard
    4. algorithmic registration