A new arXiv preprint addresses how to quantify uncertainty in spatiotemporal foundation models used for Earth system modeling. The authors frame the problem around decision-making and risk control rather than pure prediction accuracy. The work aims to make such models more reliable when their outputs inform real-world choices.
A new arXiv preprint introduces a method for autonomous agents to weigh sensory modalities against each other when inputs are noisy, incomplete, or contradictory. The approach infers how reliable each stream is and uses that estimate to arbitrate between modalities rather than assuming all sensors are equally trustworthy. The work targets robustness in real-world environments where sensory data quality varies over time.
A revised arXiv paper presents a deep evidential regression approach for estimating forest height using multimodal satellite imagery. The method targets applications including carbon accounting, biodiversity monitoring, and ecosystem management. It aims to give accurate predictions while quantifying uncertainty in sparse-data settings.
A new arXiv preprint addresses node classification on evolving graphs, where classifiers must generalize to newly arriving nodes despite distribution shift while also providing calibrated uncertainty. The authors propose an online Bayesian approach aimed at inductive settings where safety-sensitive applications require trustworthy confidence estimates. The work targets a gap left by standard graph neural networks, which typically assume static graphs and offer limited uncertainty quantification.
A new arXiv paper presents a hybrid mathematical and machine-learning approach for irrigation decision support that also quantifies confidence in its soil-moisture forecasts. The authors note that irrigation is typically scheduled reactively, even though agriculture uses about 70% of global freshwater withdrawals. The model aims to give growers both a forward-looking moisture prediction and a measure of how reliable that prediction is.
A revised arXiv paper proposes a noise-adaptive approach to conformal inference for classification. Standard conformal methods give prediction sets with guaranteed coverage, but the work addresses how that reliability holds when label or feature noise is present. The method targets marginal coverage guarantees while adjusting to noisy data conditions.
A new arXiv paper addresses a common problem in medical image classification: transitional categories whose features overlap with neighboring classes, creating fuzzy decision boundaries. The proposed approach, adaptive conformal redistribution, aims to sharpen uncertainty-aware prediction sets so they remain informative when a case sits between two diagnoses. Conformal prediction normally provides coverage guarantees for such sets, and the work adapts that framework to inter-class transitions.
Researchers examine whether uncertainty maps reliably indicate where a sparse-view CT reconstruction is erroneous, and find a discrepancy between whole-volume scoring and localization of errors within the imaged object. They derive analytic, clamp-aware moments to produce calibrated uncertainty estimates for radiative Gaussian splatting reconstructions. The work suggests standard variance-based uncertainty is not equivalent to an error map in this setting.
Researchers present an approach that embeds knowledge of the Helmholtz operator into Gaussian process models so they can reconstruct complex-valued wavefields, including the dissipative case where the squared wavenumber becomes complex. The method is tested first on synthetic benchmarks and then applied to in vivo brain elastography, where it infers tissue properties from limited, noisy measurements. The work emphasizes uncertainty quantification alongside field estimation.
A new arXiv preprint introduces ActMap, a method that estimates how much an LLM answer should be trusted using internal activation maps captured during a single generation pass. The authors position it as an alternative to approaches that require multiple sampled outputs or rely solely on output-token probabilities. The work targets practical uncertainty quantification for large language models.
A research paper addresses how to decide which LLM-judged outputs actually need human expert review. It separates the judge's uncertainty into aleatoric uncertainty, which reflects genuine disagreement among experts and cannot be reduced by more labels, and epistemic uncertainty, which signals where expert annotation would help. The goal is to spend limited expert labeling effort on the cases where it is most useful.
Researchers present an uncertainty-aware workflow that fuses geophysical inversion results with drilling data via Bayesian deep learning to model copper mineralization in three dimensions. Applied to the Kogodai prospect in the Rudny Altai region, the method quantifies prediction uncertainty to help prioritize exploration drilling in structurally complex terrain where sampling is sparse.
Researchers have introduced ProbPlug, a plugin uncertainty network designed to attach to large language models and produce more trustworthy confidence scores for binary classification tasks. The work addresses the gap between strong LLM predictive performance and the reliability required for deployment in high-stakes settings. The paper is available as a preprint on arXiv.
A new arXiv paper frames novel-view synthesis with 3D Gaussian Splatting as a structured regression problem and applies conformal prediction to it. The goal is to give rendered views formal statistical coverage guarantees, going beyond uncertainty heatmaps that offer no such certification. The approach is aimed at settings where rendered outputs must meet verifiable reliability standards.
A new machine learning paper tackles the gap between training conditions and real-world deployment for sensor-based AI, where devices, personnel, and time periods all differ from the original training setup. The authors propose a staged evaluation methodology that quantifies uncertainty and captures performance degradation that conventional random train-test splits can hide. The work draws on data collected from multiple devices and subjects over nearly three years in an underground environment.