@DanKornas: Turn a model output into an inspectable attribution graph instead of guessing which internal features contributed. circ…
Summary
circuit-tracer is a Python library that helps interpretability researchers map model computations by calculating direct effects between transcoder features and output logits, then visualizing them as an interactive attribution graph.
View Cached Full Text
Cached at: 07/31/26, 12:45 AM
Turn a model output into an inspectable attribution graph instead of guessing which internal features contributed.
circuit-tracer is a Python library for interpretability researchers and ML builders who want to examine circuits built from pretrained transcoders.
It helps you map model computations by calculating direct effects between transcoder features, error nodes, input tokens, and output logits, then turning those results into an attribution graph.
Key features: • Circuit attribution – computes feature-to-feature and feature-to-logit direct effects. • Interactive visualization – serves graphs locally for browsing, pinning, grouping, and annotation. • Feature interventions – sets transcoder features to chosen values so you can observe output changes. • Multiple workflows – runs through Neuronpedia, Python or Jupyter, and a command-line interface. • Practical demos – includes tutorials for Gemma 2, Llama 3.2, attribution targets, and interventions.
It’s open-source (MIT license).
Link in the reply
Similar Articles
Transcoders Trace Visual Grounding and Hallucinations in Vision-Language Models
This paper presents a function-centric framework using Transcoders to trace computational pathways in vision-language models, demonstrating stronger attribution of visual grounding and the ability to predict hallucinations via graph-based features.
Prune, Interpret, Evaluate: A Cross-Layer Transcoder-Native Framework for Efficient Circuit Discovery via Feature Attribution
Researchers introduce PIE, a CLT-native framework for efficient circuit discovery via feature attribution-based pruning, achieving ~40× compression in feature selection while maintaining behavioral fidelity on IOI and Doc-String tasks.
Tokengeist: Multi-Turn Attribution Tracing in Agentic Conversations
Introduces Tokengeist, a method for tracing attribution across multiple turns in agentic conversations, revealing that flat attribution methods fail and that recursive dependency graphs are needed, achieving 90% source recall on a new benchmark (MTCABench) of 3,845 target spans.
LLMs Can Annotate Attribution Graphs
The paper presents a simple pipeline that uses LLMs to automatically group features into supernodes in attribution graphs, matching human annotator interpretability and recovering intermediate hop supernodes in a two-hop task.
@yoheinakajima: and here's what a coding agent on activegraph looks like basically you always get a trace and a graph automatically
Yohei Nakajima shares a demo of a coding agent on activegraph that automatically produces a trace and graph.