SPEX: Scaling Feature Interaction Explanations for LLMs

TL;DR

SPEX is a scalable, model-agnostic explanation method that efficiently identifies feature interactions in large language models, outperforming existing methods on long-input datasets and aligning with human annotations.

Contribution

SPEX introduces a novel sparse Fourier transform-based approach to scale interaction attributions for LLMs to large input lengths, leveraging natural data sparsity.

Findings

SPEX outperforms marginal attribution methods by up to 20% on large inputs.

SPEX accurately identifies key features and interactions influencing model outputs.

SPEX's explanations align with human annotations and reveal reasoning patterns in LLMs.

Abstract

Large language models (LLMs) have revolutionized machine learning due to their ability to capture complex interactions between input features. Popular post-hoc explanation methods like SHAP provide marginal feature attributions, while their extensions to interaction importances only scale to small input lengths ($\approx 20$). We propose Spectral Explainer (SPEX), a model-agnostic interaction attribution algorithm that efficiently scales to large input lengths ($\approx 1000)$. SPEX exploits underlying natural sparsity among interactions -- common in real-world data -- and applies a sparse Fourier transform using a channel decoding algorithm to efficiently identify important interactions. We perform experiments across three difficult long-context datasets that require LLMs to utilize interactions between inputs to complete the task. For large inputs, SPEX outperforms marginal attribution methods by up to 20% in terms of faithfully reconstructing LLM outputs. Further, SPEX successfully identifies key features and interactions that strongly influence model output. For one of our datasets, HotpotQA, SPEX provides interactions that align with human annotations. Finally, we use our model-agnostic approach to generate explanations to demonstrate abstract reasoning in closed-source LLMs (GPT-4o mini) and compositional reasoning in vision-language models.