anonymous-agp-review/attribution-graph-probing-anon

Concept Swap Explorer -- Cross-Domain Circuit Steering

Interactive visualization of circuit steering experiments across 5 knowledge domains. Explore how feature-level interventions on attribution-graph circuits redirect model outputs from one concept to another.

Domains

Use the Run dropdown in the header to switch between domains and experiment types (labeled swaps, random-feature controls, field-add controls).

What This Shows

For each domain the demo visualizes experiments where we:

1. 1.Trace the attribution-graph circuit a language model uses to answer a factual prompt
2. 2.Use Probe Prompting to identify concept-related CLT features and group them into supernodes
3. 3.Swap source features out (ablate) while amplifying target features, then measure how the model's output changes

The Steering Matrix

Each cell shows the result of swapping a source entity's circuit into a target entity's context:

Click Any Cell

See detailed results including:

• —Default vs steered model outputs
• —Token probability changes
• —Logit-flip trajectory (when available)
• —Links to Neuronpedia circuit visualizations

Experiment Types

Each domain includes multiple run types:

• —Labeled -- swap matched supernodes between source and target circuits
• —Random -- swap the same number of randomly chosen features (control)
• —Field-add -- inject only target features without ablating source (control)

Related Research

• —Circuit Tracer
• —Attribution Graphs (Anthropic)
• —Neuronpedia

Data

This Space visualizes 33,387 steering runs across 5 domains and 3 experimental conditions (labeled, random-control, field-add-control), all performed on Gemma-2-2B with Cross-Layer Transcoders.

Version: 2.0.0 Model: Gemma-2-2B License: GPL-3.0