Reproduce “Understanding AI from the Inside: The State of Mechanistic Interpretability for AI Safety” 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
Fazl Barez, Tung-Yu Wu, Hunar Batra, Gabriele Dominici, Florent Draye, Yanai Elazar, Damiano Fornasiere, Antía García and 19 more
The survey organizes mechanistic interpretability (MI) methods and proposes five dimensions for evaluating their claims. It distinguishes output testing and detectable signals from causal explanations and reliable control. In a trained sleeper-agent backdoor setting, a linear probe detected triggered behavior with over 99% AUROC; this does not show that MI makes systems safer or generalizes beyond that trigger.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.