Edit concept-linked neurons, run baseline vs intervened, and inspect evidence. This UI is a researcher-focused shell over live CB-LLM backend responses.
Search, sort, and select concepts for batch edits.
Edit activation and weight interventions for selected concepts.
Additive adds Δ, override sets a fixed value, scale multiplies the activation.
Zero removes the concept’s influence; scale dampens or amplifies its effect downstream.
| Concept | Activation | Weight | Contribution |
|---|
Visualizes how concept activations flow into predicted classes.
We reproduce and extend Concept Bottleneck LLM work by building tooling to select concept-linked neurons and apply activation interventions (ablation / amplification) while measuring downstream performance and behavioral steering.
Team: Christian Guerra, Neil Dandekar