Loscar: Location-Scale RESIT

Loscar extends RESIT (regression with subsequent independence test, Peters et al. 2014) to location-scale noise models (LSNMs), using in-context quantile regression with TabPFN. This repository contains the loscar package, the data generation and the experiments.

Project Structure

Loscar/
├── src/loscar/             # the `loscar` package (Loscar, Loscar_bivariate, etc.)
├── data/                   # data generation (dgp.py), dataset loaders and the datasets (not in git)
├── experiments/            # benchmark scripts
├── tutorial.py             # Loscar on a simulated graph and a cause-effect pair
├── environment.yaml        # conda environment for Loscar
├── setupR.R                # R packages for the SID metric
├── LICENSE.txt             # license of this repository
└── LICENSES/               # licenses of adapted code, data and TabPFN

Installation & Quick-Start

The following contains instructions to install the loscar environment, as well as the packages required to run the experiments and compute the performance metrics. To test the environment and get an example for running Loscar, please have a look at the tutorial.py file. Note that for all code that uses TabPFN, the corresponding licenses provided by Prior Labs GmbH need to be accepted.

1. Create the environment

The file environment.yaml contains everything needed for the tutorial and the Loscar experiments, including the R tool chain for the SID metric.

$ conda env create -f environment.yaml
$ conda activate loscar-env

2. Install Loscar

From the repository root, install the loscar package:

$ pip install -e .

3. Install the R packages for SID

The SID metric is computed in R via cdt. Install the R packages pcalg and SID with:

$ Rscript setupR.R

Without R, the experiments still run; the SID columns are then reported as NaN. cdt prints warnings about missing R packages when imported, which can be ignored if you do not need the SID metric.

4. Run the tutorial

Run it from the repository root (it imports data.dgp). On the first run, TabPFN downloads its model weights and asks you to accept their license in the browser (see TabPFN license).

$ python tutorial.py

Datasets

The datasets are not part of the repository. Download them into data/ as described below; data/datasets.py and data/real_world.py expect exactly these folder names.

• Cause-effect pairs: All pairs are taken from the data folder of the skewd repository. These include datasets from Mooij et al. (2016), Goudet et al. (2018), Tagasovska et al. (2020), and Klippert and Marx (2026).
• SACHS (Sachs et al., 2005): observational data available at the BCD-Nets repository.
• LightTunnel (Causal Chambers, Gamella et al., 2025): downloaded automatically to data/causal_chambers/ on first use via the causalchamber package. The data is licensed under CC BY 4.0, see LICENSES/causal-chambers-4.0.txt.
• Synthetic graphs: All synthetic graphs are generated on the fly with data/dgp.py.

Running the Experiments

The scripts to reproduce the experiments involving all Loscar and RESIT variants are provided in the experiments folder.

License and Third Party Code

The code in this repository is licensed under the MIT License, see LICENSE.txt. Code and data adapted from other projects keep their original licenses, see LICENSES/ and the notes at the top of the respective files. In particular, this project builds on the following resources:

• https://github.com/cdt15/lingam for the base RESIT algorithm and the HSIC implementation.
• https://github.com/causality-lab/skewd for the bivariate experiments pipeline.
• https://github.com/xunzheng/notears as a basis for the synthetic multivariate experiments.

TabPFN license

Loscar uses TabPFN by Prior Labs GmbH for in-context quantile regression. TabPFN is not included in this repository. It is installed as a dependency and its weights are downloaded on first use.

• TabPFN package (code): Apache License 2.0, see LICENSES/tabpfn-Apache-2.0.txt.
• TabPFN-3 weights (the default, tabpfn-v3-regressor-v3_default.ckpt): TABPFN-3 Non-Commercial License, see LICENSES/tabpfn-3-weights-license.txt or the original.
• TabPFN-3.5 weights (optional, --model-version v35): TABPFN-3.5 Non-Commercial License, see LICENSES/tabpfn-3.5-weights-license.txt or the original.

The weight licenses allow only non-commercial use (e.g. academic research and benchmarking); this also applies to their outputs. Every TabPFN-based method in this repository (Loscar, Loscar-GL, RESIT-TabPFN, Loscar_bivariate) is therefore restricted to non-commercial use, unless you obtain a commercial license from Prior Labs. Loscar-FGLS and RESIT-spline do not use TabPFN.

Citation

@article{klippert:26:loscar,
  author = {
    Klippert, Daniel and
    Huynh, Quang Nhat and
    Marx, Alexander
  },
  title = {Residual-Based Causal Discovery via Pretrained In-Context Quantile Regression},
  journal={arXiv preprint arXiv:2610.29530},
  year = {2026}
}

References

• Gamella, Peters, and Bühlmann, P. Causal chambers as a real-world physical testbed for ai methodology. Nature Machine Intelligence, 2025
• Goudet et al. Learning functional causal models with generative neural networks. Explainable and Interpretable Models in Computer Vision and Machine Learning, 2018
• Klippert and Marx. Skewness-robust causal discovery in location-scale noise models. ICML, 2026
• Mooij et al. Distinguishing cause from effect using observational data: Methods and benchmarks. JMLR, 2016
• Peters et al. Causal discovery with continuous additive noise models. JMLR, 2014 Sachs et al. Causal protein-signaling networks derived from multiparameter single-cell data. Science, 2005
• Tagasovska, Chavez-Demoulin and Vatter. Distinguishing cause from effect using quantiles: Bivariate quantile causal discovery. ICML, 2020