A reduced phase space approach shows a lower bound on position uncertainty in systems with broken time-translation symmetry, quantizing classically branched dynamics without explicit path tracking. The method encodes momentum-position relations in a geometric framework, offering potential insights and needing further validation against standard branch-matching quantization.
A framework called multichoice modeling explicitly represents large, interdependent decision spaces under uncertainty and analyzes them with variance-based and conditional distribution methods. Case studies show it reveals complementarities and alternative risk profiles not visible in traditional uncertainty-only analyses.
URW introduces uncertainty-aware rubric weighting to account for varying certainty across evaluation criteria, using conformal calibration to assign per-criterion weights and improving alignment in downstream reinforcement learning without additional training.
Statistical uncertainties in quantum chemistry calculations on near-term quantum devices can be far smaller than actual errors; Fraunhofer Institute finds discrepancies up to 32x between reported uncertainties and true residuals, even with ZNE and TREX mitigation, underscoring the need for rigorous validation before applying results to complex problems.
The work reframes input perturbation for QA uncertainty as a quality-diversity search problem and introduces neighborhood elites to distribute informative rewrites across edit types and support explicitness, delivering superior failure detection across multiple datasets and models.
UtoMe adds observation-uncertainty guidance to token merging in weather transformers, improving forecast accuracy under constrained token budgets by combining dynamics saliency with ensemble uncertainty as a reliability proxy; three training/inference variants show gains at moderate token reductions with some losses at aggressive compression.
Translated title. Original: “Mensen vertrouwen een AI-model blindelings”: waarom elke meting een foutmarge nodig heeft
AI models are increasingly used to interpret measurements, but their outputs often lack the margin of error that metrology requires. Researchers at Dutch institute VSL are studying how to estimate uncertainty from the model, its training data, and each new measurement.
LEASH is a training-free framework for uncertainty-aware escalation of computer-use agents, triggering a strong model only when the weak model shows uncertainty and maintaining oversight afterward. It uses latent uncertainty across three layers to route, escalate, and monitor, achieving cost-efficient performance improvements over weak models and on-par accuracy with a strong model on OSWorld benchmarks.
The paper presents an uncertainty-aware staffing framework that translates probabilistic ICU risk predictions into hourly, cost-conscious staffing decisions, using CVaR to handle shortage risk and evaluated on MIMIC-IV data. It demonstrates improved cost-shortage trade-offs and robustness over baselines, offering a principled method to integrate probabilistic AI outputs into healthcare operations.
The article explains how nnetsauce’s QuantileRegressor turns compatible regression models into prediction-interval estimators, with five scoring strategies and matching Python and R interfaces. It shows a diabetes-data example reporting 94.7% coverage against a 95% target and outlines a broader benchmark across models, datasets, and coverage levels.