method
active
method:linear-steeringlinear steering
Typical approach that adds a scaled steering vector to representations; the paper argues this is mismatched with actual representation geometry.
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (2)
finding
- Central empirical result showing causal coupling between representation and behavior geometry across multiple substrates and modalities.
- Empirical demonstration on Llama-3.1-8B that steering along representation manifold aligns outputs with behavior manifold, whereas linear steering does not.
Concepts (7)
concept
- Manifold SteeringcontradictsextendsCentral framework: steering neural networks by intervening along the curved manifold where a concept lives, rather than in straight lines through activation space.
- steering vectorsimplementsA method for modifying model behavior by adding perturbation vectors to activations, used here to try to reduce eval awareness.
- Euclidean Geometry AssumptionimplementsThe implicit assumption of linear steering methods, which the paper argues is inappropriate for neural activation spaces
- Linear steering implicitly assumes a flat, Euclidean activation space, leading to off-manifold excursions.
- Representation SteeringextendsParent concept; the practice of controlling neural network outputs by manipulating internal representations.
- direction-based steeringimplementsParadigm of finding the right direction in activation space (e.g., linear steering).
- steering (intervention on internals)implementsGeneral technique of modifying activations to control model behavior.
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.
- The paper's critique of the standard linear steering baseline, supported by the days-of-week demo.
- A straight vector in activation space, traditionally used for concept manipulation; claimed to be insufficient when true concept geometry is curved.
- Attribution of failure to Euclidean assumption.
- Steering variants that adaptively modulate steering strength based on input context or token position, addressing when to steer
- Constructing steering vectors from the difference of mean activations on positive and negative examples, for comparison.
- Manipulation of activations along a straight line; shown to fail when it crosses voids, in contrast to manifold-following interventions.
- Novel method that applies intervention only when the model begins a new thinking step (at the \n\n delimiter) rather than at every token
- Steering variant that rotates hidden states toward a target feature vector, addressing how to steer