Reproduce “Multi-component Causal Tracing in 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
Zirui Yan, Dennis Wei, Dmitriy A. Katz, Prasanna Sattigeri, Ali Tajer
PGB-CT uses gradient-optimized soft masks to select sparse sets of language-model components for causal tracing, enabling joint interventions without exhaustive subset search. It patches counterfactual states and optimizes a chosen metric. On GPT2-medium WinoGender at 40% sparsity, its score is 0.233, near greedy’s 0.237, with runtime of 1.56 versus 357.28 minutes; results are limited to tested settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.