Reproduce “Decision-Aligned Evaluation of Uncertainty Quantification” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Annika Schneider, Tommy Rochussen, Joshua Stiller, Vincent Fortuin

Researchers introduce a framework to assess the decision-alignment of Uncertainty Quantification (UQ) metrics, revealing that common evaluation methods often misalign with real-world decision utility or encode problematic prior beliefs. The paper proposes Prior-Weighted Utility (PWU) metrics, which are empirically shown to consistently align with actual downstream decision utility across various decision problems and real-world applications.

Key insights

  • Common UQ metrics like NLL, Brier Score, and Accuracy are often decision-aligned only under 'pathological' implicit priors, such as assuming unbounded mass on extreme costs or perfectly symmetric costs.
  • Coordinate-dependent metrics (e.g., ECE, R-AUC) are found to not be decision-aligned for common pointwise-separable utility functions, indicating they cannot consistently rank models based on decision utility in these contexts.
  • No conventional UQ metrics adequately address decision problems with non-separable utilities, such as top-k selection, as they evaluate the entire dataset rather than the relevant subset.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.