2026· Proceedings of the 21st International Conference on Software Technologies· pp. 360-368· 0 citations· 18 references
TL;DR
A trust-centric framework that integrates ensemble-based uncertainty estimation, maximum mean discrepancy (MMD) shift detection, a dependability signal, and safety-aware decision logic to prioritize reliable actions over raw predictive confidence is proposed.
Abstract
: Accuracy alone is insufficient for machine learning systems that support high-stakes decisions under uncertainty and distribution shift. This paper proposes a trust-centric framework that integrates ensemble-based uncertainty estimation, maximum mean discrepancy (MMD) shift detection, a dependability signal, and safety-aware decision logic to prioritize reliable actions over raw predictive confidence. The framework first estimates prediction reliability and then modulates action selection through feasibility constraints and a conservative fall-back policy when dependability is low. Evaluation on three high-stakes decision scenarios under controlled distribution-shift stress tests shows that the proposed framework reduces high-confidence errors from 18.3% to 6.8%, decreases constraint violations from 12.7% to 1.4%, and improves decision stability from 0.42 to 0.22 relative to an accuracy-only baseline, while maintaining competitive average utility. Ablation results further show that uncertainty weighting, shift detection, and fallback control each contribute to these gains. These findings support the need for reliability-aware, constraint-aware decision systems in safety-critical applications.
Predictive Decision Support Systems (PDSS) increasingly rely on machine learning, but conventional models often provide deterministic predictions without measuring uncertainty, limiting their reliability in high-stakes applications. This paper proposes an uncertainty-aware machine learning framework that integrates uncertainty quantification, probabilistic modeling, explainable AI, and adaptive learning to improve prediction reliability, transparency, and decision confidence. The framework models both aleatoric and epistemic uncertainties, producing confidence scores, prediction intervals, and interpretable decision reports instead of single-point predictions. By combining intelligent data preprocessing, uncertainty-aware learning, trustworthy decision intelligence, and continuous model adaptation, the proposed approach enhances robustness, calibration, explainability, and trustworthiness, providing a scalable foundation for reliable predictive decision support in dynamic real-world environments.
David Wheeler, Michael Gordon· International Journal of Mac...· 0 citations
Large Language Model (LLM) ensembles are increasingly used to improve reliability by combining predictions from multiple LLMs. However, existing aggregation methods typically assume that all models are equally trustworthy, overlooking differences in uncertainty quality. This assumption is poorly suited to heterogeneous LLMs, whose reliability and capability vary significantly, making naive aggregation vulnerable to unreliable or adversarial experts. In this work, we formulate multi-LLM aggregation as a problem of uncertainty-aware trust estimation. We adapt structured expert judgment from decision theory, using context-aware calibration questions to estimate expert reliability based on the quality of its probabilistic predictions. Specifically, we employ Cooke-style log weighting, which penalises overconfident incorrect predictions and favours well-calibrated experts. We evaluate our approach on MMLU and MMLU-Pro across homogeneous, heterogeneous, and contaminated expert panels. Results show that while aggregation methods perform similarly in homogeneous settings, Cooke weighting becomes critical under heterogeneity and contamination. It achieves a superior accuracy-reliability balance and remains robust when unreliable experts are introduced. These findings suggest that Multi-LLM aggregation requires not just combining predictions, but calibrating trust under uncertainty.
This has cast extreme doubt in relation to transparency, accountability, and human trust in machine learning applications despite the growing dependency on automated decision-making systems. Numerous high-performance models are black boxes, limiting human knowledge and supervision in high-stakes areas. This study comes up with a human-centered machine learning infrastructure that will underpin transparent automated decision-making by instantiating explainability infrastructure and human-in-the-loop architectures. The paper compares four machine learning models, namely, Logistic Regression, Decision Tree, Random Forest, and XGBoost with SHAP explanations, on a structured dataset (10,000 instances). As the experimental results indicate, XGBoost with SHAP model has best predictive performances with an accuracy of 89.2, an F1-score of 0.88 and an AUC of 0.92, with excellent explanation clarity. Random Forest had an accuracy of 86.9% and Decision Tree and Logistic Regression had accuracy of 82.4 and 79.6 respectively, but it had better interpretability. Human-in-the-loop assessment also revealed higher rates of 91.6 and lower rates of 8.4 in decision acceptance of explainable models and override, respectively. The proposed architecture has a better balance between the performance and transparency as compared to the existing related work. The results validate the claim that explainability and human supervision increase the levels of trust, usability, and ethical adherence in the automated decision-making systems
M. Manimaran, Sridhar D, Manikandan K B et al.· 2026 6th International Confe...· 0 citations
Reliability set estimation identifies input regions where a response probability exceeds a target level, bridging estimation and safety-critical decisions. Practitioners typically start with a working model, an imperfect approximation of the true response surface. Relying on this imperfect model may incur decision risk, potentially certifying unsafe regions as safe. We develop a unified framework that turns such a working model into a trustworthy decision rule. First, a modeling-then-calibration procedure decouples estimation from decision. Since the true set is unobservable, we introduce an asymmetric, observable surrogate loss and use a separate calibration set to select a bias-correcting threshold, reducing decision risk and achieving $O_P(1/n)$ volume convergence. Second, we leverage conformal risk control with the surrogate loss to control false inclusion risk, which is the most safety-critical error, at a pre-specified level regardless of working model quality. Together, these calibration procedures show that a separate calibration set is necessary for risk control. Third, an adaptive design concentrates observations on the reliability set and its boundary, improving model quality where errors most affect decisions while controlling budget elsewhere. Numerical studies show not only more accurate set estimates but also calibrated finite-sample risk control that classical plug-in methods lack.
To address the fundamental challenge of translating probabilistic health assessments into reliable maintenance decisions for elevator mechatronic systems, this paper proposes the first uncertainty-aware dual-threshold optimization framework that systematically bridges Bayesian perception with optimal decision-making. Leveraging posterior health distributions (mean and variance) output by the Bayesian Fusion for Elevator Health Assessment, the method constructs supervised learning pairs from multi-year historical records. A three-objective composite loss function-integrating matching accuracy, uncertainty coverage, and decision separability-is optimized via Bayesian optimization to calibrate optimal thresholds. Critically, the 95% confidence interval is explicitly embedded as a constraint, ensuring high-uncertainty equipment remains safely distant from decision boundaries. Validated on one-year operational data from 100 elevators, the framework achieves 84.0% matching accuracy (17.0% improvement), 0.71 maintenance cost index, and 0.12 safety risk index, delivering Pareto-optimal trade-offs among decision accuracy, maintenance economy, and operational safety. The offline learning and online deployment paradigm requires no additional hardware and integrates directly into existing supervision platforms.
Zhenkuo Kang, Qicai Zhou, Yao Wang et al.· 2026 IEEE International Conf...· 0 citations
Safety classifiers deployed with large language models often fail for two reasons: their decisions reflect the policy learned during training rather than the deployer's desired policy, and their performance degrades as deployment traffic evolves. We present Regime-Conditional Verification (RCV), a lightweight wrapper that adapts an off-the-shelf safety classifier without retraining it. RCV estimates, from the classifier's internal representations, the probability that each prediction disagrees with the deployer's policy, and selectively corrects predictions likely to be wrong. The same correctness estimates also provide a label-free signal for detecting distribution shift, enabling a maintenance loop that updates the correctness estimation layer and resorts to classifier fine-tuning only when necessary. Across three off-the-shelf safety classifiers and two benchmark datasets, RCV improves adherence to the deployer's policy in every classifier-dataset combination, catching up to 0.81 of previously missed unsafe content without modifying the underlying classifier. In a deployment study with ten attack campaigns, each a harm category held out of RCV's training, RCV detects every campaign in a dedicated injection panel; in the maintenance census most drift episodes are repaired without updating the classifier, and the fine-tune is reserved for the residual episodes that repair does not restore.