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..
DeCaFlow: A deconfounding causal generative model
Authors: Alejandro Almodóvar, Adrián Javaloy, Juan Parras, Santiago Zazo, Isabel Valera
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our empirical results on diverse settings including the Ecoli70 dataset, with 3 independent hidden confounders, tens of observed variables and hundreds of causal queries show that De Ca Flow outperforms existing approaches, while demonstrating its out-of-the-box applicability to any given causal graph. In this section, we assess the performance of De Ca Flow relative to existing methods. Namely, we show that De Ca Flow accurately estimates interventional and counterfactual queries when the requirements of Corols. 3.2 and 4.3 are met. All experimental details are described in B. |
| Researcher Affiliation | Academia | Alejandro Almodovar1, * Adrian Javaloy2,* Juan Parras1 Santiago Zazo1 Isabel Valera3 1Information Processing and Telecommunications Center, Universidad Politecnica de Madrid, ES 2School of Informatics, University of Edinburgh, UK 3 Department of Computer science, Saarland University, DE |
| Pseudocode | Yes | We additionally, provide algorithms to help practitioners easily check in the given the causal graph whether a particular query of interest can be correctly estimated by De Ca Flow. Appendix F: Algorithms for causal query identification, Algorithm 1 KL regularization term in the training loop, Algorithm 6 Identification of causal queries that include intervention and outcome (t, y), Algorithm 7 Identification of causal queries, intervening in t and evaluating in all variables |
| Open Source Code | Yes | An implementation of De Ca Flow can be found in github.com/aalmodovares/De Ca Flow. |
| Open Datasets | Yes | Our empirical results on diverse settings including the Ecoli70 dataset [66]... with the protein-signaling network dataset [63]. ...on the law school dataset [78], which comprises of 21 790 law students. |
| Dataset Splits | Yes | Both experiments were performed with 25,000 data, split into 80%, 10%, 10% (train, validation, and test). All metrics are given over the test dataset, and hyperparameter search was performed over the validation dataset. |
| Hardware Specification | Yes | All the experiments were conducted on CPU. Although the experiments were carried out on a cluster of different CPU, we include here two tables for the two semi-synthetic datasets (Tab B.6 and Tab B.7) with the processing times measured in a CPU Intel(R) Core(TM) i7-13650HX laptop, just to show that even in a laptop CPU, the training and inference times are sensible even for large datasets as the Ecoli70 dataset. |
| Software Dependencies | No | The paper mentions types of components like 'Adam optimizer [32]' and 'neural spline flow (NSF) [14]' and cites relevant papers, but does not provide specific version numbers for general software dependencies or programming languages (e.g., Python, PyTorch) or for the mentioned libraries themselves. |
| Experiment Setup | Yes | Hyperparameter selection. We choose hyperparameters based on the MMD [21] over validation observational data, following our theoretical premise that De Ca Flow correctly estimates causal queries when p M(x) = pθ(x); see B.6.3 for details on the hyperparameter grid. Although including all hyperparameters would be very extensive, we give here a sample of the hyperparameters selected for De Ca Flow in the Ecoli70 additive dataset: Hidden neurons of causal flow (generative network): 3 128 Type of causal flow (generative network): neural spline flow (NSF) [14]. Hidden neurons of encoder flow (inference network): 3 64 Type of normalizing flow (inference network): neural spline flow (NSF) [14]. Regularize: True (warm-up: 30 epochs) Total number of parameters: 182k. |