New VPD Method Decomposes Model Parameters | dailyai.report
23 stories from today
Safety
114d ago
New VPD Method Decomposes Model Parameters
The adVersarial Parameter Decomposition (VPD) method allows researchers to decompose attention layers, a known weakness for transcoders and SAEs. This technique builds attribution graphs using causally important parameter subcomponents. The authors claim the approach is now ready for scale.
The Signal
It offers a more granular path toward interpreting how specific weights influence model outputs.