AI Paper Atlas

Research map / Label-efficient and robust learning / OOD and anomaly detection

EigenScore: OOD Detection using Posterior Covariance in Diffusion Models

Shirin Shoushtari, Yi Wang, Xiao Shi, Salman Asif, Ulugbek S. Kamilov

ICLR 2026 · Label-efficient and robust learning › OOD and anomaly detection · approach: Spurious-aware OOD detection

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
replace NLL / score-norm metrics with posterior-covariance eigenvaluesdiffusion OOD detection · Average over 12 InD-OOD pairsAUROC: 0.817 (DDPM-OOD) -> 0.869Improved
use top-K eigenvalues instead of scalar MSE (ablation)EigenScore · 12 InD-OOD pairsAUROC: 0.652 -> 0.834Improved
use EigenScore instead of MSE across pairs (ablation)EigenScore · SVHN InD pairs vs C10/C100/SVHNAUROC: MSE 0.994 vs EigenScore 0.990 (SVHN vs C10); C10 vs SVHN 0.184 -> 0.778Mixed
increase number of timesteps T from 5 to 10 (ablation)EigenScore · 12 pairs averageAUROC: 0.834 -> 0.808Hurt
increase repetitions I from 5 to 20 (ablation)EigenScore · 12 pairs averageAUROC: 0.832 -> 0.834No clear effect
use K=1 instead of K=3 eigenvalues (ablation)EigenScore · 12 pairs averageAUROC: 0.834 -> 0.840Mixed
compare near-OOD C10 vs C100 against DiffPathEigenScore · C10 vs C100AUROC: 0.554 -> 0.884Improved
compare to DiffPath on TinyImageNet OODEigenScore · C10 vs TinyImageNetAUROC: 0.993 -> 0.973Hurt

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