Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
ProxySPEX: Inference-Efficient Interpretability via Sparse Feature Interactions in LLMs
Authors: Landon Butler, Abhineet Agarwal, Justin Kang, Yigit Efe Erginbas, Bin Yu, Kannan Ramchandran
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 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. |
| Researcher Affiliation | Academia | Landon Butler Department of EECS UC Berkeley EMAIL Abhineet Agarwal Department of Statistics UC Berkeley EMAIL Justin Singh Kang Department of EECS UC Berkeley EMAIL Yigit Efe Erginbas Department of EECS UC Berkeley EMAIL Bin Yu Departments of Statistics and EECS UC Berkeley EMAIL Kannan Ramchandran Department of EECS UC Berkeley EMAIL |
| Pseudocode | Yes | Algorithm 1 Fourier Extraction from Gradient Boosted Trees [5] Require: Gradient boosted model M Ensure: Fourier mapping F 1: Initialize F 2: for Tree T in M do 3: F F.merge(EXTRACTTREE(T.root)) Add mappings of the individual trees 4: end for 5: return F 6: procedure EXTRACTTREE(node n) 7: if n is leaf then 8: return { 7 n.value} 9: else 10: NL EXTRACTTREE(n.left Child) 11: NR EXTRACTTREE(n.right Child) 12: N 13: for S in (NL.keys NR.keys) do 14: v L NL[S] Mapping returns 0 if not contained 15: v R NR[S] 16: N[S] (v L + v R)/2 17: N[S {n.feature Split}] (v L v R)/2 18: end for 19: end if 20: return N 21: end procedure |
| Open Source Code | Yes | The PROXYSPEX algorithm is available at https://github.com/mmschlk/shapiq. |
| Open Datasets | Yes | 1. Sentiment is a classification task composed of the Large Movie Review Dataset [46] which consists of positive and negative IMDb movie reviews. 2. Hotpot QA [49] is a generative question-answering task over Wikipedia articles. 3. Discrete Reasoning Over Paragraphs (DROP) [50] is a paragraph level question-answering task. 4. MS-COCO [42] contains images and corresponding text captions. ... On CIFAR-10 [6] PROXYSPEX identifies the interactions between training samples that most significantly impact classification performance. |
| Dataset Splits | Yes | Step 2 Proxy Training. Fit GBTs to D with 5-fold cross-validation (CV). For each sample with n features, we generate α n log2(n) masks, varying α {2, 4, 6, 8}, to normalize difficulty across inputs of varying lengths (some by over 100 tokens). MMLU includes 231 questions in the high-school-us-history subset. To perform pruning and then evaluate the ablated models, we split this data into two sets training split Dtrain consisting of the first 120 questions and test split Dtest with the remaining questions. |
| Hardware Specification | Yes | All experiments are run on a single NVIDIA H100 GPU, with batch size 50. |
| Software Dependencies | No | For PROXYSPEX, a list of GBT hyper-parameters we tune over are in Appendix B. We use the LASSO implementation from scikit-learn, and choose the l1 regularization parameter via 5-fold CV. ... we utilize a Distil BERT model [47] fine-tuned for sentiment analysis [48]. ... We use a Llama-3.2-3B-Instruct model with 8-bit quantization. ... We use CLIP-Vi T-B/32, a joint vision-language encoder... The paper mentions software tools and models but does not provide specific version numbers for general software dependencies like Python, PyTorch, or scikit-learn. |
| Experiment Setup | Yes | We performed 5-fold cross-validation over the following hyper-parameters for each of the models: Model Hyper-parameter LASSO L1 Reg. Param. λ (100 with λmin/λmax = 0.001) SPEX L1 Reg. Param. λ (100 with λmin/λmax = 0.001) PROXYSPEX Max. Tree Depth [3, 5, None] Number of Trees [500, 1000, 5000] Learning Rate [0.01, 0.1] L1 Reg. Param. λ (100 with λmin/λmax = 0.001) Random Forest Max. Tree Depth [3, 5, None] Number of Trees [100, 500, 1000, 5000] Neural Network Hidden Layer Sizes [( n / 4 )] Learning Rate [Constant, Adaptive] Learning Rate Init. [0.001, 0.01, 0.1] Number of Trees [100, 500, 1000, 5000] For each sample with n features, we generate α n log2(n) masks, varying α {2, 4, 6, 8}, to normalize difficulty across inputs of varying lengths (some by over 100 tokens). |