acceptodds
Under review as a conference paper at ICLR 2027

Decision-Aligned Evaluation of Uncertainty Quantification

Abstract

Uncertainty estimates in machine learning are typically evaluated using generic metrics such as the negative log-likelihood and expected calibration error, yet good performance on such metrics does not necessarily imply high utility in downstream decisions. We introduce *decision-alignment*, a criterion that reveals which metrics faithfully rank models according to their decision-making performance. Applying this framework, we show that many widely used uncertainty evaluation metrics are either misaligned with common decision problems or encode implausible prior beliefs about the downstream task. We then propose *prior-weighted utility metrics*, a special class of proper scoring rules that provides decision-aligned uncertainty evaluation. Across benchmark experiments and real-world case studies, our metrics consistently align with realized decision utility, while conventional metrics do not. Our results surface gaps in the current UQ evaluation protocol and offer a principled evaluation scheme toward decision-relevant UQ evaluation.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
evaluation metrics, uncertainty quantification, uncertainty evaluation, decision-making, proper scoring rules, decision-utility

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_mmosOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper asks whether the standard metrics used to benchmark uncertainty quantification (NLL, Brier score, accuracy, ECE/MCE, retention-curve AUC, error detection) rank models by their usefulness for downstream decisions. It defines decision-alignment: a metric is aligned with a parametric family of utilities if it is a strictly increasing transform of the expected utility under some prior. This is shown to be equivalent to preserving the order and ties of that expected utility.

The authors audit 32 metric–utility pairs: binary cost-sensitive decisions, top-k selection, selective prediction, one-vs-rest multiclass decisions and multivariate selective prediction. They conclude that every metric is either not aligned or aligned only under an "implausible" implicit prior.

As a remedy they propose prior-weighted utility (PWU) metrics. The expected utility under an explicitly chosen (Beta) prior. These are aligned and proper by construction. Experiments on 40 datasets report that PWUs give the highest rank agreement with downstream utility and the lowest regret when the top-ranked model is deployed.

Critical strengths

Paper turns integral representation of a proper scoring rule into a general criterion. Metric alignment as a property of a triple of metric, utility function and a prior on utility parameters.

This work also applies this principle in a large case study, for non-separable utilities, top-k selection, selective prediction in regression, and multi class and multivariate tasks.

Experimental section is also quite substantial, conducting a lot of experiments, highlighting the strength of proposed method.

Critical weaknesses

To the best of my understanding the tackled problem has been already largely studied in the literature. The binary-decision metric PWU‑BD coincides, up to the constant with the Beta-weighted proper scoring rule of Buja et al. . Reading NLL, Brier score and accuracy as expected cost under implicit cost distributions is established (Schervish 1989; Hernández‑Orallo et al. 2012).

Buja et al - Loss Functions for Binary Class Probability Estimation and Classification: Structure and Applications, https://www.semanticscholar.org/paper/Loss-Functions-for-Binary-Class-Probability-and-and-Buja-Stuetzle/d6706b6e626c15680688b0774419662f2341caee

Schrevish 1989 - A General Method for Comparing Probability Assessors - https://projecteuclid.org/journals/annals-of-statistics/volume-17/issue-4/A-General-Method-for-Comparing-Probability-Assessors/10.1214/aos/1176347398.full

Jos´e Hern´andez-Orallo et al - A Unified View of Performance Metrics: Translating Threshold Choice into Expected Classification Loss - https://www.jmlr.org/papers/volume13/hernandez-orallo12a/hernandez-orallo12a.pdf

Without this contribution, in my opinion, theoretical results stay limited.

Questions for authors

My main concern is with novelty of the proposition, it looks like the question has been already looked at in the literature. Could you help me position your work better in what have already been done. Maybe you could highlight the part of your work that can clarify the positioning exactly.

Apart from this I appreciate the paper’s goal of making UQ evaluation decision-aware, and I think the framework is a useful lens. I would like to understand better how a practitioner or benchmark designer should use it in the general-purpose setting the paper targets, where no specific downstream decision is known.

For general-purpose benchmarking, where no downstream decision is known, which decision families and priors would you recommend as a default PWU panel, and why?

If PWUs for different decision families rank methods differently, how should a benchmark summarise or interpret the results?

With no deployment context, a uninformative prior ( i.e. uniform prior ) reduces PWU‑BD to the Brier score. What do PWUs add in this case, and when is an asymmetric default such as Beta(2,10) justified?

Can the framework evaluate uncertainty scores, such as epistemic vs. total uncertainty or mutual information, and not only full predictive distributions?

In general evaluation one would not have any information about the downstream decision. You show experimntally that even in that case your PWU metric is aligned with decision utility, could a theoretical result of that flavor be obtained ? For example showing that PWU is better aligned with some unknown PWU instantiation than classical metrics ?

Reviewer_XQcNOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

The authors introduce a novel framework for uncertainty estimates evaluation. The work comes from the observation that, despite popular uncertainty estimates metrics, such as Expected Calibration Error and Brier's Score, do not align with the downstream for which uncertainty quantification is employed for a model. The authors define an alignment measure and show, using the top-k selection task, how most of these metrics are not aligned to this task, i.e., they evaluate uncertainty well in isolation, but they are providing poor estimates with reference to the downstream task of selecting the top-k predictions with the highest utility. Finally, the authors introduce Prior Weighted Utility, a family of uncertainty estimates evaluation metrics that align to the downstream task by construction.

Critical strengths
  • The paper addresses a fundamental issue of uncertainty estimation which can lead to practical concerns whereas some misaligned evaluation metrics are applied
  • The formalization seems solid and well substantiated
  • The authors propose a new family of metrics in addition to showing the failure points of the existing ones
Critical weaknesses

I do not think I have enough understanding of the theoretical framework behind this paper to express well-substantiated critical points on it. My evaluation on the readabilty stems from this.

I can only point out an improvement point for the paper: the authors mainly motivate/exemplify their framework to the top-k selection, but, at least, according to my expertise in trustworthy AI, the reason the uncertainty layer is introduced in an AI system is often to defer uncertain predictions to human operators (or to discard the prediction), turning a binary classification into a ternary classification problem (discard/defer/keep) depending on the confidence value. I think the paper could be even better if the authors included references on how their framework can accommodate this problem formulation.

Questions for authors

None

Reviewer_pr4NOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces decision-alignment, a framework for assessing whether uncertainty quantification metrics rank predictive models consistently with their usefulness for downstream decisions. It analyzes existing metrics, including negative log-likelihood, Brier score, and accuracy, revealing their implicit assumptions about decision costs and demonstrating limitations for certain decision problems. To address these limitations, the authors propose Prior-Weighted Utility (PWU) metrics, which integrate downstream decision utility over a prior distribution of possible costs, constraints, and preferences, and establish their decision-alignment and properness under suitable assumptions. Experiments across classification, regression, and applied decision-making scenarios suggest that PWU metrics better reflect downstream decision quality than conventional uncertainty evaluation metrics.

Critical strengths
  1. The paper highlights an important limitation of conventional uncertainty evaluation: metrics that assess probabilistic accuracy or calibration do not necessarily reflect the usefulness of predictions for downstream decisions.
  2. The introduction of decision-alignment provides a principled way to analyze evaluation metrics in relation to downstream utilities. The theoretical characterization and impossibility results offer useful insights into the limitations of existing metrics.
  3. PWU provides a unified construction applicable to different decision problems, including binary classification, top-k selection, and selective prediction. The framework explicitly accounts for uncertainty about downstream costs, constraints, and preferences.
  4. The authors examine the proposed metrics across several prediction settings and include applied decision-making scenarios.
Critical weaknesses
  1. Several foundational definitions are insufficiently emphasized in the main text. In particular, the relationship between action utilities, Bayes-optimal actions, and prediction-induced utilities could be presented more explicitly. The concrete PWU instantiations also require combining definitions and assumptions from multiple sections and appendices.
  2. PWU is explicitly constructed by integrating downstream utility, making its alignment with that same integrated utility immediate. Consequently, some experimental improvements over conventional uncertainty metrics are expected. And even then it still is on par with NLL in two out of 3 presented examples in fig 4.
  3. PWU requires selecting a prior over decision parameters, which can substantially affect model rankings. Although the authors provide practical guidelines and investigate prior misspecification, the proposed procedure remains largely heuristic.
Questions for authors
  1. Could the authors clarify the practical implications of properness without strict properness, particularly for PWU-TS, which cannot distinguish probabilistic predictions that induce identical rankings?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear reviewers,

thank you for your assessment of our work and engagement in the rebuttal. In response to your feedback, we made the following main changes to our submission: We

  • improved clarity by stating relevant definitions more prominently (Rem. 3.1, Sect. 4.1, App. H),
  • improved PWU usage guidelines by adding a concrete PWU panel recommendation (Sect. 6), and advising against using non-strictly proper PWUs in isolation (below Prop. 4.2),
  • added a three-way decision (predict, defer, reject; App. F.1.1), which increases the variety of our decision examples,
  • transferred our framework to uncertainty scores (App. A), which extends our scope beyond first-order predictive distributions.

We appreciate your helpful feedback and your time for reviewing our paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.