Do Active SAE Feature Planes Carry More Holonomy? A Preregistered Reversal in Gemma

Short summary

This preregistered study tests whether holonomy concentrates on active SAE feature planes in Gemma 2 2B, operationalizing a semantic-concentration prediction from interpretability research. The prediction was falsified in reverse: active-feature planes carried less holonomy than matched mixed-feature controls, with an adjusted log contrast of -0.294. The result is a narrow, auditable operational reversal rather than a causal claim, with multiple alternative mechanisms remaining open including activation-strength geometry and transport shear.

• •Preregistered study found active SAE feature planes in Gemma 2 2B carry less holonomy than mixed-feature controls, reversing the prediction
• •Result is an auditable operational reversal, not a causal claim that meaning suppresses holonomy
• •Multiple alternative mechanisms remain open including activation geometry, dictionary geometry, and transport shear

Generated with AI, which can make mistakes.