ConceptJacobian Lens
Jacobian Lens
A method that compresses the Jacobian matrix of a Transformer's intermediate hidden states with respect to output logits into low-dimensional coordinates indexed by vocabulary, representing the sensitivity profile of intermediate states to vocabulary token predictions.
Timeline (last 90 days)
Oct 26
The Jacobian lens compresses the matrix of partial derivatives of an intermediate-layer hidden state with respect to the final output logits into low-dimensional coordinates indexed by vocabulary, decoupling the communication signal from model width.
Unverified50%