SPEX: Scaling Feature Interaction Explanations for LLMs
This work addresses the bottleneck in explainable AI for LLMs by enabling scalable interaction attributions, which is incremental as it builds on existing explanation methods but extends them to long-context datasets.
The authors tackled the problem of scaling feature interaction explanations for large language models (LLMs) by proposing SPEX, a model-agnostic algorithm that efficiently handles inputs up to 1000 features, outperforming marginal attribution methods by up to 20% in faithfully reconstructing LLM outputs.
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.