Why single measurements miss what's in neural network weights

When models are shared without documentation, identifying their architecture or training properties from weights alone is hard. Existing lightweight probing methods extract features from weight matrices but only capture first-order patterns. MVProbe adds Gram-matrix views to capture row-column interactions, balancing contributions across perspectives via scaling laws. Beats ProbeX on the Model Jungle benchmark across ResNets, vision transformers, and generative adapters.