Skip to main content
grasp.study
About
Explore
Pricing
Log in
Try it for free
Back to course
๐
Welcome to a Grasp lesson
Create a free account
to keep studying and create your own study path.
Mechanistic Interpretability and Causal Testing
ยท
Module 4
Circuits: Testing Explanations of Computation
1
Ranking Components with Gradient-Based Attribution Patching
Use gradient-based attribution patching to rank candidate components for subsequent exact intervention tests.
2
Testing Transformer Component Connections with Edge-Specific Activation Patching
Test a hypothesized connection between two transformer components using edge-specific activation patching.
3
Comparing Individual and Joint Ablations for Redundancy and Non-Additive Effects
Compare individual and joint component ablations to detect redundancy or non-additive effects.
4
Evaluating Circuit Faithfulness on Held-Out Examples
Measure a candidate circuit's behavioral faithfulness to the full model on held-out examples.
5
Testing Circuit Completeness Through Excluded Component Effects
Test a circuit's completeness by checking whether excluded components retain an effect when the candidate circuit is present.
6
Exploiting Inactive Pathways Through Activation Intervention
Demonstrate in a toy network how a successful activation intervention can exploit a pathway inactive in the unmodified model.
Previous module
Activation Interventions and Causal Controls
Next module
Sparse Autoencoders: Interpretable Features and Their Failure Modes