Jasper

Conformer Ensembles Selectively Improve Solvation-Dependent Predictions but Cannot Match End-to-End 3D GNNs

3D conformer ensemble approaches using Distribution Kernel Operators show statistically significant improvements (8-13% RMSE reduction) only on solvation-dependent molecular properties, while providing no benefit for electronic or steric tasks. This selectivity is mechanistically grounded in improved information density of conformer mean features (2-8x per feature vs. fingerprints), yet covariance features contribute <2% of model signal. End-to-end 3D GNNs substantially outperform all conformer ensemble methods (21-42% improvement over fingerprints), establishing a clear hierarchy where pre-computed feature approaches cannot compete due to inherent information bottlenecks.

Adaptive Rerouting Amplifies Cascading Losses from Maritime Chokepoint Disruptions

Maritime chokepoint closures create dynamic, non-linear economic losses that static network analysis cannot predict. An agent-based model of 35,954 ships across 1,651 ports reveals that adaptive rerouting initially reduces losses at directly exposed ports but triggers cascading delays that propagate through subsequent vessel cycles and dependent regions. Closure duration becomes a critical multiplier—each additional day reduces global shipping arrivals by 3.0% (Suez) to 7.7% (triple closure)—and uncertainty about disruption end-dates materially worsens outcomes, positioning chokepoint risk as a temporal and behavioral problem rather than a static topological one.

GP-GOMEA with GPU-Based Fitness Evaluations: Design and Performance Analysis

Learning from Saturated Data: Signals Beyond Correctness for LLM Training

Closed-Form Lower Expectations Under TV Distance and Chi-Squared Divergence Uncertainty Sets

De Bock (2026) derives closed-form expressions for lower expectations over uncertainty sets defined by total variation (TV) distance and chi-squared divergence balls around a probability mass function on a finite set. These results belong to the imprecise probability / robust Bayesian framework, where one computes worst-case expected values over a neighborhood of distributions rather than a single point estimate. Closed-form solutions are practically significant because lower expectations under such divergence balls typically require solving constrained optimization problems; analytic expressions eliminate that computational burden and enable direct integration into downstream algorithms.

The Singer-Zauner gap for equiangular tight frames

An $L^{\#}$ Based Algorithm for Active Learning of Minimal Separating Automata

Minerva-Ego: Spatiotemporal Hints for Egocentric Video Understanding

Random spanning tree Markov random field priors for Bayesian inverse problems in imaging

Dramaturgies of Deception: AI Humanizers and the Performance of Legitimacy in Higher Education Assessment

Not All Proofs Are Equal: Evaluating LLM Proof Quality Beyond Correctness

$H$-linear magnetoresistance in NbSe$_2$ due to impeded cyclotron motion