LLMs Can Annotate Attribution Graphs

Ameen Patel, Max Zhang, Nathan Hu

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…