j-lens — reading the internal geometry
Patolli asks from outside. J-lens reads from inside.
A probe can only ever see what a model does. The Jacobian lens looks at what a model is about to do, and where that comes from: take an internal activation — the residual stream at any layer, any position — and decode what that activation is disposed to make the model say, in the model’s own vocabulary.
How it works, in plain terms
Deep inside a transformer, every token position carries a vector — a bundle of numbers that will, many layers later, become the model’s next word. The lens transports that vector forward to the final layer in one linear step (the average input–output Jacobian over a text corpus), then decodes it with the model’s own unembedding. The result is a ranked list of words: this is what this spot in the machine is leaning toward saying.
That turns the invisible middle of the network into something you can read — layer by layer, position by position. Where does a preference first become legible? Does “mercy vs punishment” exist at layer 10, or only at layer 30? When a model changes its answer under a disguise, where inside did the change happen?
This is the inside half of the question Patolli measures from outside — and it runs entirely on our own hardware, on open-weight models, which is what makes the lens-grinding cheap enough to do at scale.
Status — early work
This section is being stood up. The lens itself is real and published — our reference implementation accompanies Verbalizable Representations Form a Global Workspace in Language Models — and the lab tree behind this page is active (multi-hop sweeps, lens–mirror fusion experiments, a scaling campaign across model families). Pages land here as pieces clear the same bar as the rest of the site: legible, reproducible, honestly framed.