Technical Summary: The acute Protocol

1. Problem Statement

As Large Language Models (LLMs) are increasingly deployed for critical tasks such as information seeking, writing, and agentic tool-calling, the ability to determine when to trust a model output becomes essential. While confidence estimators are used to adjudicate this trust, standard metrics like Expected Calibration Error (ECE) and its smooth variant (smECE) are insufficient for decision-making contexts.

The paper identifies two major flaws in current calibration metrics:

1. Indistinguishability from Uninformative Estimators: ECE and smECE cannot distinguish between a perfect oracle (which knows the correct answer) and a "baserate" estimator (which blindly predicts the base rate of correctness for all inputs). Both can achieve an ECE of 0, yet the baserate estimator provides no utility for decision-making.
2. Risk Invariance: These metrics are invariant to the risk level of a task. A confidence estimator that performs perfectly on a low-risk task (where a 0.5 probability threshold is sufficient) but fails on a high-risk task (requiring 0.9 probability) receives the same score as one that fails the low-risk task but succeeds on the high-risk one.

Furthermore, existing methods often rely on raw token probabilities, which are known to be poorly calibrated, or require expensive post-hoc recalibration that may not scale efficiently.

2. Methodology

The paper introduces two core components: a new evaluation metric (euro) and a new confidence estimation protocol (acute).

2.1. The Metric: Expected Utility Renormalized by the Oracle (euro)

To address the limitations of ECE, the authors propose euro, a metric that balances calibration with decision-making utility.

• Decision Framework: The metric is grounded in Minimum Bayes Risk (MBR) decision theory. It defines four outcomes based on a confidence threshold $\tau$: True Positives (trust correct), False Positives (trust incorrect), False Negatives (abstain on correct), and True Negatives (abstain on incorrect).
• Utility Definition: Each outcome is assigned a reward ($R_{tp}, R_{tn}, R_{fp}, R_{fn}$). The authors reparameterize these into two intuitive quantities:
  • $U_{ct}$: Net utility of correctly trusting a correct generation.
  • $U_{ca}$: Net utility of correctly abstaining from trusting an incorrect generation.
• Risk Parameterization: The threshold $\tau$ is defined as the normalized net utility of correctly abstaining ($u_{ca}$), allowing the metric to explicitly account for task risk levels (low, medium, high).
• Normalization: The utility of a confidence estimator is normalized against an Oracle policy (perfect prediction) and an Anti-Oracle policy (perfectly wrong prediction).
• Final Score: The metric is summarized as auc-euro, the area under the curve of the normalized utility across all possible risk settings ($u_{ca} \in [0, 1]$). A higher auc-euro indicates a better balance of calibration and utility.

2.2. The Protocol: Activation-Based Confidence Estimation (acute)

The acute protocol leverages the internal activations of the LLM to train a lightweight confidence estimator.

• Input Features: Instead of relying solely on raw token probabilities, acute extracts features from the model's hidden states during generation. Three feature types are explored:
  1. Mean-pooled Activations: Raw activations collapsed across the sequence at specific layers (early, middle, late).
  2. Layer-wise Cosine Similarities: The cosine similarity between the pooled activation of each layer and the final layer.
  3. PCA Transformed Activations: Principal Component Analysis applied to layer activations to reduce dimensionality while retaining variance.
• Classifier: These features are fed into a Random Forest classifier trained to predict whether a generated output is correct or incorrect.
• Efficiency: The protocol is designed to be compute- and sample-efficient. The inference overhead of the trained Random Forest is negligible (e.g., ~0.00007 seconds per example) compared to the generation time.

3. Key Contributions

1. The euro Metric: A new, interpretable metric that resolves the inability of ECE to distinguish between informative and uninformative estimators and incorporates task-specific risk levels.
2. The acute Protocol: A general-purpose, activation-based method for estimating confidence, utility, and trust. It utilizes model internals to create estimators that are sample-efficient and computationally lightweight.
3. Empirical Validation: Extensive experiments across 6 LLMs from 4 families (including Gemma, Qwen, Phi, and SmolLM) on three distinct tasks:
   • Multiple Choice Question Answering (MMLU).
   • Tool-calling (APIGen).
   • Scientific Document Summarization (SCITLDR).

4. Results

The paper reports results averaged across models and tasks, highlighting the following findings:

• Superior Utility: The acute estimators consistently outperform strong baselines (Raw Confidence, Histogram Regression Estimator, Nadaraya-Watson Kernel Regression) on auc-euro. This indicates that acute provides better decision-making utility, particularly in high-risk settings where baselines tend to fail or bias toward the base rate.
• Calibration Maintenance: While improving utility, acute maintains low calibration error (smECE), comparable to or slightly better than the best baselines.
• Sample Efficiency: acute methods outperform baselines even with very small training sets (e.g., 25 examples), whereas baselines like NWKR require larger datasets to achieve similar performance.
• Layer Sensitivity: Analysis of layer-wise performance reveals that confidence signals are weak in early layers, increase linearly through middle layers, and plateau in late layers. Middle and late layers generally yield the best estimators.
• Generalization: The protocol generalizes well to larger Mixture-of-Experts (MoE) models (e.g., Qwen3-30B-A3B), suggesting the approach scales effectively.
• Post-hoc Calibration: Applying post-hoc calibration (Platt scaling, Isotonic regression) to acute outputs generally reduces calibration error further without degrading auc-euro, though the gains are model and task-dependent.

5. Significance and Claims

The paper claims that equipping LLMs with the acute protocol significantly improves the reliability of confidence estimates, thereby enhancing trustworthiness in diverse settings. By introducing euro, the authors provide a more reliable framework for evaluating confidence estimators that aligns with the practical needs of risk-aware decision-making.

The authors emphasize that while acute requires access to model internals, this is a feasible requirement for model developers and organizations deploying these systems. They position their work as a step toward better metrics and methods for adjudicating trust, noting that their decision-theoretic framework is a complement to, not a replacement for, careful data analysis and system design. The work aims to address the critical gap between model performance and the ability to safely utilize that performance in real-world applications.