# ProxySPEX: Inference-Efficient Interpretability via Sparse Feature Interactions in LLMs

arXiv: https://arxiv.org/abs/2505.17495 · HTML: https://mlsanity.com/paper/2505.17495/

## Key facts

- Cited by 16 papers in the map (data through 14 Oct 2026).
- Builds on 45 of its 66 references that are in the map.
- First posted on arXiv 23 May 2025.

## Builds on (most cited first, 30 of 45)

- [Microsoft COCO: Common Objects in Context](https://mlsanity.com/paper/1405.0312/) (2014)
- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [A Unified Approach to Interpreting Model Predictions](https://mlsanity.com/paper/1705.07874/) (2017)
- ["Why Should I Trust You?": Explaining the Predictions of Any Classifier](https://mlsanity.com/paper/1602.04938/) (2016)
- [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://mlsanity.com/paper/1910.01108/) (2019)
- [HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering](https://mlsanity.com/paper/1809.09600/) (2018)
- [Axiomatic Attribution for Deep Networks](https://mlsanity.com/paper/1703.01365/) (2017)
- [DROP: A Reading Comprehension Benchmark Requiring Discrete Reasoning Over Paragraphs](https://mlsanity.com/paper/1903.00161/) (2019)
- [Towards Automated Circuit Discovery for Mechanistic Interpretability](https://mlsanity.com/paper/2304.14997/) (2023)
- [SemDeDup: Data-efficient learning at web-scale through semantic deduplication](https://mlsanity.com/paper/2303.09540/) (2023)
- [Learning to Explain: An Information-Theoretic Perspective on Model Interpretation](https://mlsanity.com/paper/1802.07814/) (2018)
- [Attribution Patching Outperforms Automated Circuit Discovery](https://mlsanity.com/paper/2310.10348/) (2023)
- [Layer-wise Relevance Propagation for Neural Networks with Local Renormalization Layers](https://mlsanity.com/paper/1604.00825/) (2016)
- [Deep learning generalizes because the parameter-function map is biased towards simple functions](https://mlsanity.com/paper/1805.08522/) (2018)
- [Open Problems in Mechanistic Interpretability](https://mlsanity.com/paper/2501.16496/) (2025)
- [The Shapley Taylor Interaction Index](https://mlsanity.com/paper/1902.05622/) (2019)
- [The merged-staircase property: a necessary and nearly sufficient condition for SGD learning of sparse functions on two-layer neural networks](https://mlsanity.com/paper/2202.08658/) (2022)
- [Scatterbrain: Unifying Sparse and Low-rank Attention Approximation](https://mlsanity.com/paper/2110.15343/) (2021)
- [ContextCite: Attributing Model Generation to Context](https://mlsanity.com/paper/2409.00729/) (2024)
- [Why are Sensitive Functions Hard for Transformers?](https://mlsanity.com/paper/2402.09963/) (2024)
- [Generalization on the Unseen, Logic Reasoning and Degree Curriculum](https://mlsanity.com/paper/2301.13105/) (2023)
- [Faith-Shap: The Faithful Shapley Interaction Index](https://mlsanity.com/paper/2203.00870/) (2022)
- [A Fine-Grained Spectral Perspective on Neural Networks](https://mlsanity.com/paper/1907.10599/) (2019)
- [The staircase property: How hierarchical structure can guide deep learning](https://mlsanity.com/paper/2108.10573/) (2021)
- [From Shapley Values to Generalized Additive Models and back](https://mlsanity.com/paper/2209.04012/) (2022)
- [SHAP-IQ: Unified Approximation of any-order Shapley Interactions](https://mlsanity.com/paper/2303.01179/) (2023)
- [Approximating the Shapley Value without Marginal Contributions](https://mlsanity.com/paper/2302.00736/) (2023)
- [Route Sparse Autoencoder to Interpret Large Language Models](https://mlsanity.com/paper/2503.08200/) (2025)
- [Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition](https://mlsanity.com/paper/2407.00886/) (2024)

## Built on by (most cited first, 16 of 16)

- [AutoSOTA: An End-to-End Automated Research System for State-of-the-Art AI Model Discovery](https://mlsanity.com/paper/2604.05550/) (2026)
- [Regression-adjusted Monte Carlo Estimators for Shapley Values and Probabilistic Values](https://mlsanity.com/paper/2506.11849/) (2025)
- [PolySHAP: Extending KernelSHAP with Interaction-Informed Polynomial Regression](https://mlsanity.com/paper/2601.18608/) (2026)
- [Tractable Shapley Values and Interactions via Tensor Networks](https://mlsanity.com/paper/2510.22138/) (2025)
- [Compressed Sensing for Capability Localization in Large Language Models](https://mlsanity.com/paper/2603.03335/) (2026)
- [Selecting Feature Interactions for Generalized Additive Models by Distilling Foundation Models](https://mlsanity.com/paper/2604.13332/) (2026)
- [Mechanistic Tomography: Designed Measurement for Control-Oriented Interpretability](https://mlsanity.com/paper/2608.19338/) (2026)
- [Cross-LLM Consistency in Inference: Evidence from Shared Interactions](https://mlsanity.com/paper/2606.08129/) (2026)
- [CUBE: Contrastive Understanding by Balanced Experiments](https://mlsanity.com/paper/2509.10825/) (2025)
- [SurrogateSHAP: Training-Free Contributor Attribution for Text-to-Image (T2I) Models](https://mlsanity.com/paper/2601.22276/) (2026)
- [Green Shielding: A User-Centric Approach Towards Trustworthy AI](https://mlsanity.com/paper/2604.24700/) (2026)
- [ConfoundingSHAP: Quantifying confounding strength in causal inference](https://mlsanity.com/paper/2605.10533/) (2026)
- [Reconciling Contradictory Views on the Effectiveness of SFT in LLMs: An Interaction Perspective](https://mlsanity.com/paper/2605.17967/) (2026)
- [TN-SHAP-G: Graph-Structured Tensor Network Surrogates for Shapley Values and Interactions](https://mlsanity.com/paper/2606.01540/) (2026)
- [Evaluating and Explaining Prompt Sensitivity of LLMs Using Interactions](https://mlsanity.com/paper/2608.18539/) (2026)
- [Spectral Feedback for Test-Time Alignment of Protein Diffusion Models](https://mlsanity.com/paper/2609.30456/) (2026)

## Abstract (arXiv)

Large Language Models (LLMs) have achieved remarkable performance by capturing complex interactions between input features. To identify these interactions, most existing approaches require enumerating all possible combinations of features up to a given order, causing them to scale poorly with the number of inputs n. Recently, Kang et al. (2025) proposed SPEX, an information-theoretic approach that uses interaction sparsity to scale to n ≈ 103 features. SPEX greatly improves upon prior methods but requires tens of thousands of model inferences, which can be prohibitive for large models. In this paper, we observe that LLM feature interactions are often hierarchical -- higher-order interactions are accompanied by their lower-order subsets -- which enables more efficient discovery. To exploit this hierarchy, we propose ProxySPEX, an interaction attribution algorithm that first fits gradient boosted trees to masked LLM outputs and then extracts the important interactions. Experiments across four challenging high-dimensional datasets show that ProxySPEX more faithfully reconstructs LLM outputs by 20% over marginal attribution approaches while using 10× fewer inferences than SPEX. By accounting for interactions, ProxySPEX efficiently identifies the most influential features, providing a scalable approximation of their Shapley values. Further, we apply ProxySPEX to two interpretability tasks. Data attribution, where we identify interactions among CIFAR-10 training samples that influence test predictions, and mechanistic interpretability, where we uncover interactions between attention heads, both within and across layers, on a question-answering task.