Measuring Uncertainty Calibration

Paper published in ICLR 26. Link to paper.

Authors: Kamil Ciosek, Nicolò Felicioni, Sina Ghiassian, Juan Elenter Litwin, Francesco Tonolini, David Gustafsson, Eva Garcia-Martin, Carmen Barcena Gonzalez, Raphaëlle Bertrand-Lalo

Environment setup

Requires Python >= 3.11 and uv.

1. Install prox_tv build dependencies

prox_tv has no pre-built wheels and is compiled from source during install. This requires OpenBLAS.

macOS

brew install openblas
export CPPFLAGS="-I$(brew --prefix openblas)/include"
export LDFLAGS="-L$(brew --prefix openblas)/lib"

Linux (Debian/Ubuntu)

sudo apt install build-essential libopenblas-dev liblapacke-dev

2. Install Python packages

Once the build dependencies above are in place, run uv sync to install all Python packages (including prox_tv):

uv sync

Running the notebook

Either open the notebook in an IDE (e.g. VSCode) and select .venv as the kernel, or start a Jupyter Lab server:

uv run jupyter-lab notebooks/sample_efficiency_corrections.ipynb

BibTeX

@article{
  ciosek2026measuringcalibration,
  title={Measuring Uncertainty Calibration},
  author={Kamil Ciosek and Nicol{\`o} Felicioni and Sina Ghiassian and Juan Elenter Litwin and Francesco Tonolini and David Gustafsson and Eva Garcia-Martin and Carmen Barcena Gonzalez and Raphaëlle Bertrand-Lalo},
  journal={ICLR},
  year={2026},
  url={https://arxiv.org/pdf/2512.13872},
  note={}
}