LLMs Automate Attribution Graph Annotation for Circuit Tracing

Circuit tracing is a powerful technique for understanding the internal computations of large language models (LLMs), but it typically involves a labor-intensive manual step: grouping individual features or Multi-Layer Perceptron (MLP) neurons into higher-level "supernodes." This manual annotation process can be a significant bottleneck for researchers. This research introduces a straightforward pipeline that automates this critical step. The method involves directly presenting descriptions of individual features to an LLM, which then groups them into meaningful supernodes. This leverages the LLM's own understanding and reasoning capabilities to interpret and organize the internal components of another model. Using automated interpretability metrics, the study confirms that the supernodes generated by this LLM-driven pipeline are just as interpretable as those created by human annotators. For instance, in a two-hop Capitals task, the pipeline successfully identified a supernode corresponding to the intermediate hop in 97 out of 100 prompts. The paper also provides a proof-of-concept for open-ended exploration, where the pipeline automatically annotated 1000 attribution graphs from Wikipedia prompt completions, with an LLM judge flagging interesting graphs for human review. This work highlights the potential for simple automation to produce valuable attribution graph annotations, paving the way for more efficient circuit tracing research.