From dense weights to a sparse read/write circuit A dense starting matrix fades into a residual while sparse A and B matrices and their bipartite read/write circuit become the protagonist. SPARSE WEIGHT DECOMPOSITION Dense weights become sparse read/write components Selected bottleneck units carry task behavior through sparse input and output edges. W ≈ A × B read (A) write (B) start: dense W inputs outputs inputs bottleneck units outputs

Hover or tap a bottleneck unit to trace its read → write path.