LLMs Can Annotate Attribution Graphs

Researchers at an undisclosed institution introduced a pipeline that uses a language model to automatically group features and MLP neurons into supernodes for circuit tracing, matching human annotator interpretability. On a two-hop Capitals task, the pipeline recovered the intermediate hop supernode in 97 of 100 prompts. The team also demonstrated open-ended exploration by annotating 1000 attribution graphs from Wikipedia prompt completions and using an LLM judge to flag interesting graphs for human review.

arXiv:2608.02632v1 Announce Type: new Abstract: Circuit tracing is an exciting technique for revealing the internal computation of language models, but it requires a time-intensive manual step of grouping individual features or MLP neurons into supernodes. We present a simple pipeline for automating this step: directly presenting feature descriptions to a language model that groups them into supernodes. Using automated interpretability metrics, we confirm that supernodes generated by our pipeline are as interpretable as those generated by human annotators. On a two-hop Capitals task, our pipeline recovers a supernode corresponding to the intermediate hop in 97 of 100 prompts. Finally, we present a simple proof of concept using our pipeline for open-ended exploration, where we automatically annotate 1000 attribution graphs from Wikipedia prompt completions and then use an LLM judge to flag interesting graphs worth human review. We hope this work demonstrates that even simple automation can produce meaningful attribution graph annotations, motivating further work on automated circuit tracing.