Mechanistic Tomography for AI Interpretability and Control

Mechanistic interpretability aims to uncover hidden states, component effects, and interactions within AI models that are not directly exposed. This research proposes "mechanistic tomography" as a unified framework to describe how different measurement techniques, such as patching, gradients, and subset interventions, contribute to understanding these internal mechanisms and their responses to interventions. The core idea is to treat these measurements as a system where observed responses (y) relate to target internal maps (x) through an intervention matrix (A) and noise (w). The framework offers a practical methodology: begin with the least costly measurements, validate them on held-out interventions, and refine simple mismatches. If structured residuals persist, expand the family of measurements. The paper highlights that control systems provide a rigorous validation context, as an estimate guiding an intervention acts as an observer, with control error linked to observer error. Experiments on various models, including HMMs and GPT-2-small, demonstrate the framework's utility. For instance, sparse aggregate measurements can recover finite-effect maps with fewer interventions than coordinate patching, and lifted measurements can reveal interactions missed by first-order maps. The study underscores that the appropriate measurement family depends on the chosen basis and the specific internal mechanisms being investigated.