Behavior-based Path

The path in activation space derived by optimizing steering interventions to produce outputs along the behavior manifold, independent of representation geometry.

Neighborhood — ranked by edge-count

Methods (1)

method

Pullback Steering
introduces
The method of optimizing steering interventions in activation space to produce outputs that follow the behavior manifold, independent of the representation manifold.

Concepts (2)

concept

behavior manifold
associated_with
One-dimensional curved surface in output probability space; the paper shows this mirrors representation manifold structure.
Representation-based Path
analogous_to
The path in activation space derived by fitting the representation manifold, used to steer along the geometric structure of internal representations.

Findings (1)

finding

The representation-based path and the behavior-based path in Llama-3.1 8B activation space trace out similar curves, demonstrating bidirectional geometry alignment.
cites
Key empirical result showing that optimizing for behavioral outputs and fitting representation geometry produce the same path in activation space.

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Behavioral Trajectoryconcept0.819
The path traced through output probability distribution space as interventions are applied to activations
Goal-Directed Behaviorconcept0.797
Observable behavioral pattern used to infer cognition; shared by plants and animals and proposed as evidence for sentience.
Behavior-Optimized Activation Path Recoverymethod0.794
Method of optimizing activation-space interventions to produce behavioral paths along M_y, then measuring whether the resulting activation trajectories trace M_h curvature
Adaptive Behaviorconcept0.772
Organism's belief-guided action selection that instantiates generative model and maintains phenotypic states
Behavior Spaceconcept0.766
A geometric space of all output token probability distributions, equipped with Hellinger distance, used to visualize model behavior.
Path Patchingmethod0.765
Method by Goldowsky-Dill et al. 2023 for localizing model behavior via targeted activation interventions
Path-Based Activation Interventionmethod0.763
The general experimental approach of intervening along geometrically-defined paths rather than single-point or linear-direction interventions
Epistemic Behaviorconcept0.762
World-disclosing behavior that resolves uncertainty; driven by epistemic value and novelty components of expected free energy