uncertainty-quantification

The paper proposes RUPA, a framework that models LLM agent execution as a dependency graph to propagate uncertainty, improving failure detection and confidence estimation in long trajectories.

This paper proposes Structure-Preserving Epistemic Neural Networks (S-PENNs), a framework for uncertainty quantification in scientific machine learning models with hard architectural constraints, instantiated for GENERIC dynamics to ensure thermodynamically consistent rollouts and calibrated prediction intervals with reduced computational cost.

This paper proposes a scalable local Sinkhorn divergence framework for training stochastic neural networks to reconstruct multidimensional random fields, with theoretical generalization error bounds and numerical demonstrations for uncertainty quantification.

The paper proposes A-CRC-QA, a post-hoc calibration framework for selective question answering that controls error rates among accepted answers via asymptotic risk calibration, demonstrating improved reliability-retention trade-offs on CoQA and MedMCQA.

This paper studies whether single-turn uncertainty quantification methods transfer to interactive LLM agent trajectories, evaluating white-box, black-box, and reflexive scorers across five LLMs and four tool-use datasets. Results show that transfer is uneven, with black-box self-consistency often strongest, and recommend revalidating UQ methods at the trajectory level.

Introduces Retrieval-Corrected Conformal Prediction (RCCP), a retrieval-augmented calibration method for time series prediction intervals that selects similar past residuals and applies a scalar conformal correction to achieve target coverage with low overhead.

The paper adapts classical statistical estimators to deep learning for tracing epistemic uncertainty into aleatoric and data-scarcity sources, leveraging approximate Fisher Information Matrices for scalability.

KReF introduces a training-free retrieval framework for long-term time-series forecasting that constructs empirical predictive distributions from similar historical lookback-future pairs, achieving strong CRPS performance across multiple benchmarks.

The paper introduces DUD (Decoupled Update Dynamics), a framework that separates Feed-Forward Network and Attention contributions via causal interventions to improve uncertainty quantification and calibration in large language models, outperforming state-of-the-art baselines.

This paper proposes weighted conformal methods for changepoint localization and root cause analysis that reduce confidence set size under corrupted observations by downweighting likely contaminated data, using uncertainty signals and meta-learning.

MetaKoopman proposes a Bayesian meta-learning framework for modeling nonlinear dynamics using linear latent representations via Koopman operators, enabling closed-form updates and uncertainty quantification. It is validated on autonomous truck and trailer systems under adverse winter conditions, outperforming prior methods in prediction accuracy and robustness.

Presents a novel estimation framework for D-vine copulas using gradient-based MLE and beam search for better global fit, and a localized anomaly detection method with uncertainty quantification via conformal prediction.

This paper proposes DiffDiff, a diffusion framework for probabilistic time series forecasting that embeds predictability asymmetry into the diffusion trajectory, outperforming six diffusion baselines on seven benchmarks across four prediction horizons.

Proposes the Verbalized Particle Posterior (VPP), treating verbalized learning as Bayesian inference by maintaining a population of natural-language hypotheses as particles, updated via Metropolis-Hastings or Sequential Monte Carlo, and improving over single-hypothesis VML on benchmarks.

This paper introduces Neural Atom Prevalence (NAP), a Bayesian framework for structured node-level model selection that achieves high sparsity, accuracy, and uncertainty quantification in neural networks.

ConfidenceBench is a new benchmark that evaluates verbalized confidence estimates in large language models using Brier scores, revealing that accuracy and calibration diverge and that even highly accurate models can be severely miscalibrated.

This paper presents a comparative study of Monte Carlo Dropout and Deep Ensemble methods for uncertainty quantification in AI-driven crash simulation surrogates, using an open-source bumper beam benchmark.

This paper proposes Logical Graph Uncertainty (LGU), a framework that models implication and incompatibility among answers to improve uncertainty estimation in LLMs, outperforming semantic entropy baselines by up to 7.1% AUROC and 3.5% AUARC across benchmarks.

This paper proposes a semiparametric framework for stochastic fundamental diagram modeling that combines physically-constrained functional forms with neural network structures to capture complex traffic flow patterns and uncertainty, demonstrating superior performance on real-world datasets.

This paper investigates the calibration gap that arises when temperature scaling, a common post-hoc calibration method relying on one-hot labels, is applied to models trained with soft label distributions reflecting genuine human disagreement. Experiments across vision and language domains show that temperature scaling calibrated on hard labels consistently underperforms direct soft-label calibration, with larger gaps in language tasks, highlighting risks for safety-critical deployments.