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..
SD-KDE: Score-Debiased Kernel Density Estimation
Authors: Elliot Epstein, Rajat Vadiraj Dwaraknath, Thanawat Sornwanee, John Winnicki, Jerry Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments on synthetic tasks in 1D, 2D and on MNIST, demonstrate that our proposed SD-KDE method significantly reduces the mean integrated squared error compared to the standard Silverman KDE, even with noisy estimates in the score function. |
| Researcher Affiliation | Academia | 1Stanford University, Stanford, CA 94305, USA EMAIL |
| Pseudocode | Yes | Algorithm 1 Score-Debiased Kernel Density Estimation Require: Data {xi}n i=1, score estimator ˆs, kernel K, KDE bandwidth h, score step size δ 1: Take a single step along the score function: exi xi + δˆs(xi) for i = 1, . . . , n 2: Compute the debiased kernel density estimate: ˆp(x) = 1 nhd Pn i=1 K x e xi h |
| Open Source Code | Yes | 1We open-source our implementation of SD-KDE at https://github.com/Elliotepsteino/SD-KDE. |
| Open Datasets | Yes | In this study, we follow a similar experimental setup to Liu et al. (2020) and explore the relationship between generated image quality and estimated density using the MNIST dataset a widely recognized benchmark comprising 70,000 grayscale images (28 28 pixels) of handwritten digits (Le Cun & Cortes, 2010). |
| Dataset Splits | No | The paper does not explicitly provide specific training/test/validation dataset splits. It mentions 'We sample data from three mixtures' for 1D synthetics and 'We trained a DDPM on this dataset' for MNIST, implying a training set, but without concrete percentages, sample counts, or references to predefined splits. |
| Hardware Specification | Yes | Most of the experiments in the paper were conducted on a Linux cluster with 5 NVIDIA RTX A6000 GPUs, each with 49140 MB memory, running on CUDA Version 12.5. The cluster has 256 AMD EPYC 7763 64-Core Processor CPUs. Some experiments were also conducted on a Mac Book Air (2022) equipped with an Apple M2 chip and 16 GB of unified memory. |
| Software Dependencies | Yes | Most of the experiments in the paper were conducted on a Linux cluster with 5 NVIDIA RTX A6000 GPUs, each with 49140 MB memory, running on CUDA Version 12.5. |
| Experiment Setup | Yes | For the diffusion model architecture, we use a 3-layer MLP with hidden dimension 512, and we train the model with Adam for 1500 steps. We use 1000 diffusion steps during training. ... For this experiment, we will sample 1000 points and hold the bandwidth constant at 0.15. ... score-based correction step, with scale 0.015 decaying at a rate of 0.15 at each iteration. |