Reproduce “Circuit Oracle: Automating Attribution Graph Analysis via Natural-Language Queries” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shariar Kabir, Hong Kiat Tan, Swastik Agrawal, Sai V R Chereddy, Sriram Balasubramanian
Circuit Oracle uses an off-the-shelf LLM equipped with task-specific skills to explore attribution graphs and produce explanations with pruned circuits. Across three proxy tasks, it scores 78.2% ± 2.4 on spurious-feature detection. On refusal suppression, its single committed intervention scores 0.804 versus 0.825 for a baseline; 0.854 is best-of-five, not a single pick.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.