GEMMA 2 · INTERACTIVE ACTIVATION ATLAS

Explore a prompt’s semantic neighborhood.

Perturb the Layer-10 activation, then observe the propagated response in Layer-14 SAE features and the model’s ordinary final output.

LAUNCH INSTRUMENT →READ THE DOCUMENTATION
The LLM Probe displaying a Layer-10 perturbation surface, Layer-14 SAE response and final token outputOPEN THE LIVE INSTRUMENT ↗
LAYER 10PERTURBATION LAYER

The prompt creates an anchor activation. Distance and direction move the probe within a selected two-dimensional plane.

→
LAYER 14SAE READOUT LAYER

The Layer-14 SAE reports active feature identities and magnitudes after the perturbation has propagated forward.

→
FINAL OUTPUTBEHAVIORAL READOUT

Gemma’s normal language-model head supplies token probabilities, entropy and KL divergence. No Layer-14 logit lens is used.

Measured terrain

Every node is an observed activation. Visible edges disclose which adjacent samples form each triangular face.

Three planes

Compare SAE-feature, principal-component and random directions in separate near and far fields.

Live telemetry

Follow SAE features, token displacement, feature count, entropy and divergence as the probe moves.