Safety3h ago
Safety
New VPD Method Decomposes Model Parameters
The adVersarial Parameter Decomposition (VPD) method now allows researchers to decompose attention layers, a known hurdle for SAEs and transcoders. This technique improves upon previous SPD and APD iterations. By building attribution graphs from causally important subcomponents, AI Alignment Forum researchers can now interpret internal model logic. The approach is ready for large-scale application.