Causal Probe
An improved linear probe training method that constrains learning to the causally effective low-dimensional subspace of activation space. This approach simultaneously enhances the model's steering capability and maintains concept-detection accuracy.
Timeline (last 90 days)
The paper proposes the causal probe method, which restricts the training of linear probes to the causally effective low-dimensional subspace.
Causal probes achieved a 17%–118% improvement in steering capability across different models, with only a 3% decrease in concept detection capability.
All Facts (2)
The paper proposes the causal probe method, which restricts the training of linear probes to the causally effective low-dimensional subspace.
50%UnverifiedCausal probes achieved a 17%–118% improvement in steering capability across different models, with only a 3% decrease in concept detection capability.
50%