method
active
method:intervention-experiments-inhibition-steeringIntervention experiments (inhibition/steering)
Used to validate graph-predicted mechanisms in the original model.
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (1)
concept
- Mechanistic unfaithfulnessassociated_withGraph-predicted indirect effects that do not exist in the underlying model.
Methods (1)
method
- Intervention technique clamping activations prior to the intervention layer.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Practical restriction of interventions to those producible by actual inputs; standard in DAS practice
- General technique of modifying activations to control model behavior.
- Steered story generation reveals cheating themes.
- Manipulation of activations along a straight line; shown to fail when it crosses voids, in contrast to manifold-following interventions.
- Modifying SAE hidden activations h to h-tilde while keeping error term fixed, then continuing inference.
- Demonstration of steering reasoning length.
- Intervention mode where interventions are applied sequentially, each building on the previous one
- Perturbing source features along decoder direction and measuring target change.