From Directions to Regions: Decomposing Activations in Language Models via Local Geometry

Name: From Directions to Regions: Decomposing Activations in Language Models via Local Geometry
Uploaded: 2026-10-03T23:36:16.281599+00:00
Duration: 304 s
Description: This work introduces Mixture of Factor Analyzers (MFA) to decompose language model activations into local Gaussian regions. MFA models activation space as a collection of these regions, each defined by a centroid and a low-dimensional subspace capturing local variation. This provides an interpretable view of how concepts are realized in the model's internal representations.

This work introduces Mixture of Factor Analyzers (MFA) to decompose language model activations into local Gaussian regions. MFA models activation space as a collection of these regions, each defined by a centroid and a low-dimensional subspace capturing local variation. This provides an interpretable view of how concepts are realized in the model's internal representations.

Open the full episode