Skip to main content
grasp.study
About
Explore
Pricing
Log in
Try it for free
Back to course
đź‘‹
Welcome to a Grasp lesson
Create a free account
to keep studying and create your own study path.
Mechanistic Interpretability and Causal Testing
·
Module 6
Shortcut Claims, Representation Drift, and Training Guidance
1
Estimating Shortcut Reliance with Matched Counterfactuals
Estimate shortcut reliance using matched input counterfactuals that change a nuisance cue while preserving task evidence.
2
Evaluating Shortcut Mechanisms with Counterfactuals and Internal Interventions
Evaluate a proposed shortcut mechanism by comparing input-level counterfactual results with targeted internal interventions.
3
Testing Collateral Effects of Feature Suppression on Legitimate Behavior
Test whether suppressing a shortcut-associated feature also harms legitimate behavior that uses the same representation.
4
Assessing Interpretability Signals Across Model Checkpoints
Assess a fixed interpretability signal across model checkpoints by comparing it with refreshed readouts and repeated causal tests.
5
Controlled Feature-Guided Fine-Tuning and Behavioral Evaluation
Design a controlled feature-guided fine-tuning experiment with a matched baseline and independent behavioral evaluation.
6
Auditing Shortcut Removal and Distributional Limits
Audit a claim of shortcut removal for proxy gaming, migration to unmonitored representations, and limits of the tested distribution.
End of course,
Back to course
Previous module
Sparse Autoencoders: Interpretable Features and Their Failure Modes
Next module