Work

Structural Inference: Interpreting Small Language Models with Susceptibilities

Year: 2025
Type: conference-paper
Abstract: We develop a linear response framework for interpretability that treats a neural network as a Bayesian statistical mechanical system. A small perturbation of the data distribution, for example shifting the Pile toward GitHub or legal text, induces a first-order change in the posterior expectation of an observable localized on a chosen component of the network. The resulting susceptibility can be estimated efficiently with local SGLD samples and factorizes into signed, per-token contributions that serve as attribution scores. We combine these susceptibilities into a response matrix whose low-rank structure separates functional modules such as multigram and induction heads in a 3M-parameter transformer.
Authors: Baker, Garrett, Wang, George, Hoogland, Jesse, Murfet, Daniel
Language: English
Open Access status: green
OpenAlex ID: W4415320092
FWCI: 2.902
Cites: 43
Cited by: 15
Related to: 10