Tiled SVD Extracts Mechanism Mounts from Linear Sites

A new arXiv paper (2608.06969) proposes a method for mechanistic interpretability of neural networks by extracting 'mechanism mounts' directly from linear sites using column-tiled SVD. Each mount is a triple (v, u, σ) representing trigger, write, and strength, with identity defined as the weight rule. The approach is evaluated on Gemma-2-2B with WikiText-2 (16,384-token subsample), scoring all seven linear maps. Residual writes (mlp.down, attn.o) receive full A/B/C grades with steering after post-sublayer RMSNorm and pass 52/52 site-layers; other maps (mlp.gate, attn.q, attn.k, effective mlp.up, attn.v) receive A/B grades only, passing 26/26 each. The aggregate score is 182/182 GO. The authors release library code and the corresponding dataset. This work challenges the dominant approach of training proxy dictionaries like sparse autoencoders, which label features from max-activating text, by instead locating mechanisms in the network weights themselves.