LLM-powered pipeline automates circuit tracing, recovering hops in 97% of prompts

Circuit tracing is a powerful technique for understanding what happens inside language models, but it typically requires a time-intensive manual step: grouping thousands of individual features or MLP neurons into larger 'supernodes' that represent meaningful computational units. This paper from Ameen Patel, Max Zhang, and Nathan Hu, presented at the ICML Mechanistic Interpretability Workshop 2026, introduces a simple pipeline that replaces that manual step with an LLM. Instead of humans labeling each neuron or feature, the pipeline feeds feature descriptions directly to a language model, which groups them into supernodes automatically.

Using automated interpretability metrics, the researchers show that these LLM-generated supernodes are just as interpretable as those created by human annotators. On a two-hop Capitals task, the pipeline recovered a supernode corresponding to the intermediate hop in 97 of 100 prompts, demonstrating that the automation preserves the core signal needed for circuit analysis. As a proof of concept, they applied the pipeline to open-ended exploration, automatically annotating 1,000 attribution graphs from Wikipedia prompt completions. Then, an LLM judge flagged the most interesting graphs for human review, creating a scalable triage system for interpretability researchers. This work suggests that even straightforward LLM-based automation can produce meaningful attribution graph annotations, potentially unlocking large-scale circuit tracing across many models and behaviors.