The Jacobian lens (Gurnee et al. 2026) reads a hidden activation hℓ at layer ℓ by transporting it into the final-layer basis with the corpus-averaged Jacobian Jℓ and decoding with the model’s own unembedding: lensℓ(h) = unembed(Jℓh). The tokens it surfaces are what the model is verbalizably holding at that depth — often intermediate concepts that are neither the input nor the next token.
The logit lens toggle skips the transport (unembed(h) directly) — the classic baseline. Comparing the two is the point: the J-lens recovers interpretable intermediates many layers before the logit lens does.
Row shading marks the three regions of the network: early sensory layers (readouts are noisy and largely uninterpretable — greyed out), the middle workspace band (layers 12–29 on this model — where the interpretable action is), and the final motor layers, where the readout collapses onto the actual next-token prediction. The top row is the model’s real output distribution.
Pinning a token colors every (layer × position) cell by that token’s rank in the readout (log scale: 1 · 10 · 100 · 1k+), and draws its rank trajectory in the charts — watch a concept surface in the workspace and dissolve into the motor layers.
Model: Qwen3.5-4B · lens: neuronpedia/jacobian-lens (n=1000, wikitext) · backend: Modal L40S, scales to zero when idle · prompts are logged for usage stats.