New Audit Method Clarifies AI Model Interaction Interpretations

Interpreting how different parts of a neural network interact is crucial for understanding its behavior, with order-dependent activation statistics often used as evidence of such interactions. However, this interpretation can be misleading due to where interventions are introduced into the network. A new "no-fit site-asymmetry audit" is proposed to address this issue. This audit decomposes the "open-path order-swap" into an additive response from single interventions and an antisymmetrized second difference. This second difference is specifically designed to be free of first-order and pure self-curvature terms up to the second order, providing a cleaner signal of true interaction. The study applied this audit across six open-weight language model families and other non-language models, finding that single-intervention baselines often explain a significant portion (84.3-97.7%) of the observed effects. The corrected residual, which represents genuine interaction, cleared a generic-interaction null in only a subset of the tested models. The key takeaway is to always run the single-intervention baseline before attributing order-swap vectors to interaction or geometric structure, and to instead focus on the second difference if the baseline explains the vector.