method
active
method:distance-matrix-analysis-of-concept-centroidsDistance Matrix Analysis of Concept Centroids
L2 distance between centroids compared across manifolds, RSA-like.
Neighborhood — ranked by edge-count
Papers (1)
paper
Frameworks (1)
framework
- Core proposal: LLM beliefs are distributions over a low-dimensional conceptual space; ICL is a trajectory through it.
Methods (1)
method
- Representational Similarity Analysisanalogous_toInspiration for distance-matrix comparison.
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.
- Procedure extracting concept vectors as difference of mean activations between concept-exemplifying and baseline/negative sentences
- Steering effect between concepts correlates with My centroid distance, r=.65 (p<.001) for Emotionsfinding0.727Geometry predicts steering.
- The spatial/geometric organization of conceptual structure within neural network representations; central to the paper's thesis.
- Probabilistic framework formalizing concept-specific subspaces for targeted steering in generative models.
- Novel calibration of injection strength as the distance from centroid midpoint to centroid; enables meaningful cross-layer comparison of alpha values
- Emotion centroid distances in My correlate r=.93 with human valence-arousal ratings (Russell & Mehrabian)finding0.700LLM emotion geometry matches humans.
- Computed directional vector in activation space representing a specific concept, used for injection experiments