Recasts covariance shrinkage as risk minimization over stochastic interpolants between distributions, recovering known estimators via scheduling, couplings, and early stopping, and proposing a neural estimator with quadratic risk bounds.
A well-conditioned estimator for large- dimensional covariance matrices
8 Pith papers cite this work. Polarity classification is still indexing.
citation-role summary
citation-polarity summary
years
2026 8roles
background 1polarities
unclear 1representative citing papers
Derives non-asymptotic error bounds for standard, defensive, and self-normalized importance sampling with random KDE proposals from geometrically ergodic Markov chains, separating n^{-1/2} Monte Carlo error from MIAE/MISE proposal error.
CausalHealth detects lithium-ion battery degradation with 100% sensitivity and up to 402-cycle lead time using causal anomaly scores from voltage, current, temperature, and resistance time series across seven cells.
A methodological framework for BCIs that separates speed and accuracy with Gain and Conservation measures combined into an alpha-controlled balance for tunable operating points.
Introduces interval graphical lasso to estimate a shared precision matrix for interval-valued data and proves its sparsity and consistency.
JBShield is vulnerable to adaptive JB-GCG attacks (up to 53% ASR) because jailbreak representations occupy a distinct region in refusal-direction space; the new RTV defense using Mahalanobis detection on multi-layer fingerprints reaches 0.99 AUROC and limits adaptive ASR to 7%.
Task-aligned supervised geometric stability predicts linear steerability with high accuracy while unsupervised stability detects representational drift earlier and with lower false alarms than CKA or Procrustes.
Asymmetry PRISM-CPU achieves 4.5x-24.1x speedups over reference solvers on N=100-2000 problems and GPU completes all 500 accounts in 109.5s where OSQP completes 4.
citing papers explorer
-
Revisiting JBShield: Breaking and Rebuilding Representation-Level Jailbreak Defenses
JBShield is vulnerable to adaptive JB-GCG attacks (up to 53% ASR) because jailbreak representations occupy a distinct region in refusal-direction space; the new RTV defense using Mahalanobis detection on multi-layer fingerprints reaches 0.99 AUROC and limits adaptive ASR to 7%.