Behavioral Null Space

The span of vector directions that do not change network behavior; a key concept distinguishing MAS from model stitching.

Neighborhood — ranked by edge-count

paper

framework

Model Alignment Search (MAS)
associated_with
The primary contribution of the paper: a bidirectional causal method that learns rotation matrices for each model to uncover and compare causally relevant latent subspaces across neural networks.

method

Algorithm 1: Harmlessness Classification
implements
Proposed algorithm using local PCA to classify a divergence vector as harmless or harmful via behavioral null-space testing

concept

Behavior Space
related_to
A geometric space of all output token probability distributions, equipped with Hellinger distance, used to visualize model behavior.
Harmless Divergence
associated_with
Divergences that occur in the behavioral null-space and do not affect functional claims about the model
Behaviorally Binary Subspace
associated_with
A vector subspace that causally impacts outputs only through the sign of its values, enabling harmless magnitude divergence
Dormant Subspace
associated_with
Subspace dimensions that do not vary between inputs; included in the extraneous subspace zextra.

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.

3D behavioral spaceconcept0.819
The traditional space of movement in the physical world where animals exhibit problem-solving behavior.
Physiological spaceconcept0.751
A space of physiological parameters that systems navigate for adaptive responses.
Positive Spaceconcept0.748
The property that every bit of space swells outward, is substantial in itself, and is never the leftover from an adjacent shape; every single part of space has positive shape as a center with no amorphous meaningless leftovers
Multi-dimensional delegation spaceconcept0.744
Activation spaceconcept0.739
Representation space on which linear probes operate to attribute harmful behaviors to training data.
metacognitive spaceconcept0.739
Ji-An et al.'s characterization of the limited regime in which model self-report succeeds, consistent with this paper's findings
Configuration Spaceconcept0.737
The ensemble of all possible configurations of a building, including incomplete states and paths between them.
Behavior Clusteringconcept0.732
Grouping similar model behaviors; the unsupervised method surfaces clusters of concerning patterns.