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..
ZEUS: Zero-shot Embeddings for Unsupervised Separation of Tabular Data
Authors: Patryk Marszałek, Tomasz Kuśmierczyk, Witold Wydmański, Jacek Tabor, Marek Śmieja
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 3 Experiments |
| Researcher Affiliation | Academia | 1Faculty of Mathematics and Computer Science, Jagiellonian University, Kraków, Poland 2Doctoral School of Exact and Natural Sciences, Jagiellonian University, Kraków, Poland |
| Pseudocode | No | The paper describes the methodology using mathematical formulas and descriptive text in Section 2, but no explicit pseudocode or algorithm blocks are present. |
| Open Source Code | Yes | The code used in this paper is available at https://github.com/gmum/zeus. |
| Open Datasets | Yes | For evaluation, we consider three groups of datasets: real datasets from Open ML [3] (Real), synthetic mixtures of Gaussians (Syn. Gauss.), and synthetic mixtures of Gaussians transformed by Res Net-like neural networks (Syn. Transf.). |
| Dataset Splits | Yes | Figure 4 presents two plots of average ARI over 200 synthetic validation datasets throughout 300 pre-training epochs. |
| Hardware Specification | No | Some experiments were performed on servers purchased with funds from the flagship project entitled Artificial Intelligence Computing Center Core Facility from the Digi World Priority Research Area within the Excellence Initiative Research University program at Jagiellonian University in Krakow. We also gratefully acknowledge Polish high-performance computing infrastructure PLGrid (HPC Center: ACK Cyfronet AGH) for providing computer facilities and support within computational grant no. PLG/2024/017893. |
| Software Dependencies | No | The evaluation of baseline models is based on the following libraries and Git Hub repositories: 1. scikit-learn used for k-means and GMM, 2. https://github.com/vlukiyanov/pt-dec implementation of the DEC, 3. https://github.com/dawnranger/IDEC-pytorch source code for the IDEC, 4. https://github.com/jsvir/idc official implementation of the IDC method, 5. https://github.com/mdsamad001/G-CEALS Deep-Clustering-for-Tabular-Data codebase for the GCEALS, 6. https://github.com/Prior Labs/tabpfn-extensions a library that extends Tab PFN functionality to a wide spectrum of machine learning tasks, including unsupervised ones, 7. https://github.com/clabrugere/pytorch-scarf code repository for SCARF. |
| Experiment Setup | Yes | Model architecture: ZEUS relies on a transformer architecture similar to Tab PFN [16]. It consists of 12 attention blocks, each with 6 heads and a token dimension of 512, with Ge LU activation employed. [...] Pre-training: [...] For training, we employ the Adam optimizer along with a cosine learning rate scheduler with warm-up, using a learning rate of 2e-5. [...] Inference: [...] preprocessing of each dataset involves standardizing numerical features, followed by scaling them to the range [ 1, 1], whereas categorical features are transformed using one-hot encoding. The input size of our model is fixed to 30. [...] at inference we use k-means applied to the normalized (=scaled to [ 1, 1]) transformer output in order to obtain clusters from our learned representation. |