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..

Accelerating Feature Conformal Prediction via Taylor Approximation

Authors: Zihao Tang, Boyuan Wang, Chuan Wen, Jiaye Teng

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical validations showcase that FFCP performs comparably with FCP (both outperforming the Split CP version) while achieving a significant reduction in computational time by approximately 50x in both regression and classification tasks. The code is available at https://github.com/Elvis Wang1111/Fast Feature CP. Extensive experiments with both synthetic and real data demonstrate the effectiveness of the proposed FFCP algorithm (Table 2).
Researcher Affiliation Academia Zihao Tang Boyuan Wang Chuan Wen Jiaye Teng Shanghai University of Finance and Economics Southern University of Science and Technology Shanghai Jiao Tong University Correspondence to EMAIL
Pseudocode Yes Algorithm 1 Split Conformal Prediction, Algorithm 2 Fast Feature Conformal Prediction, Algorithm 3 Fast Feature Conformalized Quantile Regression (FFCQR), Algorithm 4 Fast Feature Localized Conformal Prediction (FFLCP), Algorithm 5 Fast Feature Regularized Adaptive Prediction Sets (FFRAPS).
Open Source Code Yes The code is available at https://github.com/Elvis Wang1111/Fast Feature CP.
Open Datasets Yes Datasets. We consider both synthetic datasets and realistic datasets, including (a) synthetic dataset: Y = WX + ̈, where X [0, 1]100, Y R, ̈ N(0, 1), W is a fixed random matrix. (b) realworld unidimensional target datasets: ten datasets from UCI machine learning [Asuncion, 2007] and other sources: community and crimes (COM), Facebook comment volume variants one and two (FB1 and FB2), medical expenditure panel survey (MEPS19 21) [Cohen et al., 2009], Tennessee s student teacher achievement ratio (STAR) [Achilles et al., 2008], physicochemical properties of protein tertiary structure (BIO), blog feedback (BLOG) [Buza, 2014], and bike sharing (BIKE), (c) real-world semantic segmentation dataset: Cityscapes [Cordts et al., 2016], and (d) real-world semantic classification dataset: Imagenet-Val [Deng et al., 2009].
Dataset Splits Yes Algorithm 1 Split Conformal Prediction Input: Confidence level ̇, dataset D = {(Xi, Yi)}i I, testing point X 1: Randomly split the dataset D into a training fold Dtra {(Xi, Yi)}i Itra and a calibration fold Dcal {(Xi, Yi)}i Ical; A total of 1,200 samples were created, with 1,000 used for training and 200 for testing.
Hardware Specification Yes All the tests are performed on a desktop with an Intel Core i9-12900H CPU, NVIDIA Ge Force RTX 4090 GPU, and 32 GB memory.
Software Dependencies No The paper does not explicitly state specific version numbers for software dependencies or libraries used.
Experiment Setup No Model Architecture. For the one-dimensional we employ a four-layer neural network, with each layer consisting of 64 dimensions. For the semantic segmentation experiment, we utilize a network architecture combining Res Net50 with two additional convolutional layers. The paper does not explicitly provide specific hyperparameter values like learning rate, batch size, or optimizer settings for the training process.