Technical Summary: ECUASn – A Family of Metrics for Principled Evaluation of Uncertainty-Augmented Systems

Problem Statement

In high-stakes automated decision-making, users (human or downstream systems) require access to predictive uncertainty to decide whether to accept or reject a system's output based on application-specific cost trade-offs. These systems, termed Uncertainty-Augmented (UA) systems, output both a candidate prediction and an uncertainty score.

Current literature evaluates UA systems using disjointed metrics:

1. Prediction Quality: Measured via accuracy or error rate.
2. Uncertainty Quality: Measured via binary classification metrics (e.g., AUC, ECE, Brier score) treating uncertainty as a predictor of incorrect answers.
3. Integrated Metrics: Approaches like the Area Under the Risk-Coverage Curve (AURC) integrate these aspects by sweeping a rejection threshold.

The authors argue these approaches are inadequate. Separate metrics fail to reflect the value provided to the user making a final decision. Furthermore, AURC and its variants are invariant to monotonic transformations of the uncertainty score; they reward ranking but do not penalize uncertainties that are poorly calibrated (over- or underestimated) as long as the relative order is preserved. This is a critical flaw for systems where uncertainty must be interpreted as a probability of correctness. Additionally, existing methods often require fixing a specific rejection cost ($\gamma$) during development, which may not be known or may vary across users and instances.

Methodology

The paper proposes ECUASn (Expected Cost for Uncertainty-Augmented Systems), a family of metrics grounded in Bayesian decision theory.

1. Decision-Theoretical Foundation

The authors define a UA system as a function $f^{UA}: \mathcal{X} \to \tilde{\mathcal{D}} \times \mathbb{R}_{\ge 0}$, outputting a candidate response $\tilde{d}$ and an uncertainty score $u$. The final decision $d$ is either $\tilde{d}$ or a rejection $d_r$.

A cost function $C_\gamma(y, d)$ is defined based on a rejection cost $\gamma$:
$$C_\gamma(y, d) = \begin{cases} \tilde{C}(y, d) & \text{if } d \in \tilde{\mathcal{D}} \\ \gamma & \text{if } d = d_r \end{cases}$$
where $\tilde{C}$ is the cost of the candidate response given the ground truth $y$.

2. Derivation of Proper Scoring Rules (PSR)

To avoid fixing $\gamma$, the authors construct a metric that rewards systems producing probabilistically interpretable uncertainties.

• Bayes Decision: For a given cost function and a conditional distribution $q$ over classes, the optimal decision minimizes the expected cost. The expected cost of the optimal decision, denoted $u_{\tilde{C}}(q)$, serves as the theoretical uncertainty.
• Proper Scoring Rule (PSR): By evaluating the cost function at the Bayes decision, a new cost function $C^*_\gamma(y, q)$ is derived. This function is a Proper Scoring Rule, meaning it is minimized in expectation when the predicted distribution matches the true distribution. This property ensures that the metric rewards well-calibrated probabilities, unlike AURC.

3. The ECUASn Family

To create a metric independent of a specific rejection cost, the authors integrate the PSR $C^*_\gamma$ over a range of possible rejection costs $\gamma$, weighted by a function $w(\gamma)$.
$$C^*_w(y, q) = \int_{0}^{u_M} w(\gamma) C^*_\gamma(y, q) d\gamma$$
The authors parameterize the weight function as $w_n(\gamma) \propto \gamma^{n-1}$, resulting in the ECUASn metric. The parameter $n$ controls the trade-off:

• $n=0$: Heavily penalizes incorrect predictions made with high confidence (high uncertainty is required for rejection). This is recommended for high-stakes scenarios.
• $n \to \infty$: The metric degenerates toward the standard 0-1 error rate, becoming less sensitive to the quality of the uncertainty score itself.

4. Application to Generative Systems

For generative models, the concept of "correctness" is defined over semantic equivalence classes rather than exact string matches. The authors argue that for optimal decision-making, uncertainty should be computed over the distribution of these equivalence classes, not individual token sequences. This provides a theoretical justification for empirical findings that semantic entropy yields better uncertainty scores. In practice, for generative tasks, the number of classes $K$ is assumed to approach infinity.

Key Contributions

1. Principled Metric Replacement: Introduction of ECUASn, a family of metrics grounded in decision theory that replaces AURC. Unlike AURC, ECUASn rewards uncertainties that are probabilistically interpretable (proper scoring rules) while remaining agnostic to a specific rejection cost.
2. Theoretical Insight on Generative Uncertainty: A theoretical explanation for why uncertainty scores derived from semantic equivalence classes (rather than individual predictions) are superior for generative systems. This is derived directly from the construction of the Bayes decision cost.
3. Parameterized Trade-off: The introduction of the parameter $n$, allowing practitioners to tune the metric based on the specific risk profile of their application (e.g., $n=0$ for high-stakes rejection scenarios).

Empirical Results

The authors evaluated ECUASn on diverse classification and generation datasets:

• Classification: Experiments on language (AGNews, SST-2), speech (SITW, FVCAUS), and image (CIFAR-10/100) datasets.
  • Calibration Impact: ECUASn metrics (particularly for low $n$) successfully detected improvements in system quality due to post-hoc calibration, whereas AURC often remained unchanged because it is invariant to monotonic transformations that preserve ranking.
  • Naive Baseline: ECUASn correctly identified systems that, despite having good calibration scores (like CE or Brier score on the distribution $q$), performed worse than a naive prior-based system when considering the combined output of prediction and uncertainty.
• Generation (TriviaQA & MMLU): Experiments using various LLMs (Qwen, Ministral, Gemini) and three confidence elicitation methods (Sequence Posterior, Is True, Verbalized).
  • Metric Disagreement: The ranking of confidence elicitation methods varied significantly depending on the metric used. For instance, the "Verbalized" method often had the best error rate (ER) but poor calibration, leading to high ECUASn scores (worse performance) compared to "Sequence Posterior."
  • Ground Truth Sensitivity: The authors manually curated a subset of TriviaQA (455 samples) to correct errors in automatic semantic equivalence labeling. They demonstrated that automatic evaluation can drastically alter system rankings, highlighting the need for careful evaluation protocols.

Significance and Claims

The paper claims that ECUASn provides a comprehensive assessment of UA systems by integrating the impact of both incorrect predictions and imperfect uncertainties into a single value.

• Decision-Theoretic Alignment: The metrics are designed to reflect the actual utility of the system for a user making decisions under uncertainty, rather than just measuring statistical properties of the output.
• Calibration Sensitivity: By rewarding probabilistic interpretability, ECUASn addresses a major limitation of AURC, which fails to distinguish between well-calibrated and poorly calibrated uncertainties if the ranking is preserved.
• Practical Utility: The framework allows users to select an $n$ value appropriate for their risk tolerance, making it a flexible tool for evaluating systems in high-stakes environments where the cost of a false positive (accepting a wrong answer) is severe.

The authors conclude that while ECUASn assumes the user operates under a decision-theoretic framework (minimizing expected cost), it offers a more robust and principled alternative to current evaluation practices for uncertainty-augmented systems.