Most interpretability tools tell you which neurons fire; this one shows you where the information actually went. It treats a single forward pass as a graph — tokens flowing through attention heads and feed-forward blocks — and lets you prune that graph down to just the edges that carried signal from prompt to prediction, then click any edge to interrogate what that one head or neuron contributed.
What Sets It Apart
- Built on TransformerLens, it loads real open models (the GPT-2 and Llama families, among others) rather than toy networks — what you inspect is the architecture people actually run.
- The contribution graph is thresholded and interactive: raise the threshold and a dense model collapses to the handful of heads that mattered for a given output, so "the model did X" becomes "these three heads did X."
- Any component is queryable. Project a token's representation onto the vocabulary at any layer and read off which tokens a head or FFN block promotes versus suppresses — turning attention weights into human-readable hypotheses.
Who It's For
A strong fit if you do mechanistic-interpretability research or need to localize a specific behavior to specific heads; the click-to-drill visual workflow is faster than wiring up hooks by hand, and a hosted demo lets you try it before installing anything. Look elsewhere if you want production monitoring or to analyze closed API models — it runs locally on TransformerLens-supported models, and the repository was archived in February 2026, so treat it as a stable research artifact rather than actively maintained software.
