Reproduce “Scalable Circuit Learning for Interpreting Large Language Models” 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
Naiyu Yin, Dennis Wei, Tian Gao, Amit Dhurandhar, Karthikeyan Natesan Ramamurthy, Yue Yu
CircuitLasso infers sparse dependency maps from recorded activations, reducing the need to test candidate links one at a time. It fits ordered components with sparse regression; its coefficients represent population-level dependencies, not exact causal effects. Across 17 known-circuit tasks, it averaged 3.16 SHD and 16.3 seconds per case, versus EAP-ig’s 2.98 and 49.1 seconds. These averages are not a statistical equivalence test.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.