WHY THIS INSTRUMENT EXISTS
What happens when we nudge the tip of an activation arrow?
And how much of a nudge can a model tolerate before semantic meaning is distorted?
The rationale
A prompt produces a particular activation at a particular layer. That activation is the anchor. The LLM Probe makes controlled perturbations around it and records how the model responds—not as an abstract table, but as terrain that can be explored.
The goal is educational and empirical: to make high-dimensional response geometry tangible while keeping measured samples visibly distinct from interpolation or conjecture.
The experiment
The intervention is made at Gemma 2’s Layer 10. After the perturbed state propagates forward, the Layer-14 SAE supplies the internal feature readout. Token probabilities, entropy and KL divergence are instead calculated from Gemma’s ordinary final output after all remaining layers, final normalization and the language-model head. No Layer-14 logit lens is used.
Every result is conditioned on the selected prompt and plane. The instrument is not a map of “the model” in general; it is a collection of carefully identified slices through its activation space.
The creator
The project was created by Larry as an interactive companion to independent research on measuring perturbation response in language models, and for the talk Journey to the Centre of the LLM.
The public instrument is intended for researchers, students and curious practitioners who want to inspect the measurements directly and form their own questions.