List your company

Backdoor Detection on SimCoT-Llama-3.2-3B n=200, per-example targeting (test)

100Best AUCLinear probe
-3.89623.07750.0577.023Apr 1, 2026

Evaluation results

MethodLinks
Linear probeTier=Tier 3: Supervised (oracle labels)2026.04100100
SVD spectralTier=Tier 2: Population-contrast (clean + triggered samples)2026.0499.996.3
SAE anomalyTier=Tier 2: Population-contrast (clean + triggered samples)2026.0499.331.7
Act. ClusteringTier=Tier 2: Population-contrast (clean + triggered samples)2026.0497.887
Readout directionTier=Tier 1: Unsupervised (no trigger knowledge)2026.0447.2—
k-means clusterTier=Tier 1: Unsupervised (no trigger knowledge)2026.048.6—
Centroid distanceTier=Tier 1: Unsupervised (no trigger knowledge)2026.040.1—