Explore the structure
Start with a prompt and an output token. Navigate semantic clusters and filter the graph to focus on relevant features and connections.
From computational graphs to meaningful circuits.
Explore, inspect, and intervene in one interactive workspace.
Understanding a language model’s output requires navigating a complex network of internal features and connections. CircuitExplorer is an interactive system for exploring computational graphs in language models and finding meaningful circuits.
The workspace brings together semantic clusters, circuit graphs, and feature-level evidence. Researchers can move between an overview of the graph and individual features, inspect where features appear across layers and token positions, and compare baseline and steered outputs through intervention experiments.
Three connected perspectives on the same computational graph.
Start with a prompt and an output token. Navigate semantic clusters and filter the graph to focus on relevant features and connections.
Locate features across model layers and token positions. Examine descriptions, activations, and associations to build an interpretation.
Adjust selected features at token positions and compare steered outputs with a baseline to investigate a hypothesis.
Fact: the capital of the state containing Dallas is Austin
This walkthrough starts with a geographic question. Explore clusters related to cities, states, and capitals; inspect their features and graph connections; then test selected features with steering.
The recorded walkthrough retains “Austin” after the illustrated intervention. Comparing output text is one piece of evidence; inspecting logits and testing other intervention settings helps assess the hypothesis further.
Read the step-by-step walkthroughTry the live system, run it locally, or follow a documented exploration.