Reproduce “Demystifying Variance in Circuit Discovery of LLMs” 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
Frank Zhengqing Wu, Francesco Tonin, Volkan Cevher
The authors introduce CEAP, a conductance-based method for identifying language-model circuits, and test how stable those circuits are across samples and prompt templates. CEAP yields higher or comparable pairwise Jaccard Index (PJI) under resampling than EAP-IG, with comparable unfaithfulness. In selected models and tasks, prompt templates can produce distinct circuits, while sample-level unfaithfulness can be affected by normalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.