Max Torop

Max Torop, Aria Masoomi, Davin Hill et al.

Several recent methods for interpretability model feature interactions by looking at the Hessian of a neural network. This poses a challenge for ReLU networks, which are piecewise-linear and thus have a zero Hessian almost everywhere. We propose SmoothHess, a method of estimating second-order interactions through Stein's Lemma. In particular, we estimate the Hessian of the network convolved with a Gaussian through an efficient sampling algorithm, requiring only network gradient calls. SmoothHess is applied post-hoc, requires no modifications to the ReLU network architecture, and the extent of smoothing can be controlled explicitly. We provide a non-asymptotic bound on the sample complexity of our estimation procedure. We validate the superior ability of SmoothHess to capture interactions on benchmark datasets and a real-world medical spirometry dataset.

Davin Hill, Josh Bone, Aria Masoomi et al.

Explainability methods are often challenging to evaluate and compare. With a multitude of explainers available, practitioners must often compare and select explainers based on quantitative evaluation metrics. One particular differentiator between explainers is the diversity of explanations for a given dataset; i.e. whether all explanations are identical, unique and uniformly distributed, or somewhere between these two extremes. In this work, we define a complexity measure for explainers, globalness, which enables deeper understanding of the distribution of explanations produced by feature attribution and feature selection methods for a given dataset. We establish the axiomatic properties that any such measure should possess and prove that our proposed measure, Wasserstein Globalness, meets these criteria. We validate the utility of Wasserstein Globalness using image, tabular, and synthetic datasets, empirically showing that it both facilitates meaningful comparison between explainers and improves the selection process for explainability methods.

Max Torop, Aria Masoomi, Masih Eskandar et al.

A variety of recent methods guide large language model outputs via the inference-time addition of steering vectors to residual-stream or attention-head representations. In contrast, we propose to inject steering vectors directly into the query and value representation spaces within attention heads. We provide evidence that a greater portion of these spaces exhibit high linear discriminability of concepts --a key property motivating the use of steering vectors-- than attention head outputs. We analytically characterize the effect of our method, which we term DISentangled COmmunication (DISCO) Steering, on attention head outputs. Our analysis reveals that DISCO disentangles a strong but underutilized baseline, steering attention inputs, which implicitly modifies queries and values in a rigid manner. In contrast, DISCO's direct modulation of these components enables more granular control. We find that DISCO achieves superior performance over a number of steering vector baselines across multiple datasets on LLaMA 3.1 8B and Gemma 2 9B, with steering efficacy scoring up to 19.1% higher than the runner-up. Our results support the conclusion that the query and value spaces are powerful building blocks for steering vector methods.