AI Paper Atlas

Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes

Equivariant Covariance Tensors: Guaranteed SPD Uncertainty for Tensor-Valued Geometric Learning

Ruihan Liu, Yu Ji, Jianbo Yu, Shifu Yan, Qingchao Jiang

ICML 2026 · Probabilistic and scientific modeling › Uncertainty and Gaussian processes · approach: Priors and functional variational uncertainty

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use full covariance instead of diagonal covariance (ablation)equivariant UQ model · ModelNet40 inertiaMAE: 0.092 -> 0.078Improved
use full covariance instead of diagonal UQ (ablation)equivariant UQ model · Materials Project dielectricMAE: 2.25 -> 1.55Improved
train with LE-ESO (Laplace) instead of Gaussian NLL (ablation)matrix-exp covariance head · Materials Project dielectricMAE: 1.78 -> 1.55Improved
train with LE-ESO instead of Energy Score (ablation)matrix-exp covariance head · Materials Project dielectricMAE: 1.64 -> 1.55Improved
train with LE-ESO instead of Gaussian NLL (ablation)matrix-exp covariance head · Materials Project dielectricMACE: 0.092 -> 0.049Improved
use matrix exponential head instead of Cholesky covariance head (ablation)equivariant mean model · ModelNet40 inertiacovariance equivariance error: 4.30e-1 -> 2.75e-7Improved
use matrix exponential instead of direct equivariant regression of covariance (ablation)equivariant covariance head · ModelNet40 inertiaSPD validityImproved
apply post-hoc temperature scaling (T about 0.05) (ablation)Ours (Full Cov.) · Materials Project dielectricEnergy Score: 0.87 -> 0.66Improved
rank by lambda_max(Sigma) instead of Trace(Sigma) for risk-coverage (ablation)predicted covariance · Materials Project, 90% coverageretained-set MAE improvement: 0.8% -> 3.1%Improved
joint training without gradient detachment instead of detached UQ gradients (ablation)LE-ESO training · Materials Project dielectricMAE: MAE difference <0.02No clear effect
set LE-ESO weight alpha=1 versus smaller values (ablation)LE-ESO · validation sweepvalidation MAEImproved
add anisotropic jitter to eigenvalues for gradient stability (ablation)spectral decomposition · Materials Project dielectriccalibration (MACE): MACE change <1e-4No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy